<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Nisaharan Genhatharan on Medium]]></title>
        <description><![CDATA[Stories by Nisaharan Genhatharan on Medium]]></description>
        <link>https://medium.com/@nisaharan?source=rss-bb3486e6f2d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xYavmZs0SX4irtyFldQxPg.jpeg</url>
            <title>Stories by Nisaharan Genhatharan on Medium</title>
            <link>https://medium.com/@nisaharan?source=rss-bb3486e6f2d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 06:15:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nisaharan/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The laundered text scored more human than the human]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@nisaharan/the-laundered-text-scored-more-human-than-the-human-609c7eac8306?source=rss-bb3486e6f2d------2"><img src="https://cdn-images-1.medium.com/max/1000/0*hJFuTCdibpMfiqcJ.gif" width="1000"></a></p><p class="medium-feed-snippet">I spent a week building tools to spot machine-written text. Then I tried to disguise some, and broke my own test.</p><p class="medium-feed-link"><a href="https://medium.com/@nisaharan/the-laundered-text-scored-more-human-than-the-human-609c7eac8306?source=rss-bb3486e6f2d------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@nisaharan/the-laundered-text-scored-more-human-than-the-human-609c7eac8306?source=rss-bb3486e6f2d------2</link>
            <guid isPermaLink="false">https://medium.com/p/609c7eac8306</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[token]]></category>
            <category><![CDATA[embedding]]></category>
            <category><![CDATA[bias]]></category>
            <dc:creator><![CDATA[Nisaharan Genhatharan]]></dc:creator>
            <pubDate>Sat, 15 Aug 2026 09:43:42 GMT</pubDate>
            <atom:updated>2026-08-15T09:43:42.430Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Agents for Software Development: Experiments, Results, and Future Directions]]></title>
            <link>https://medium.com/@nisaharan/agents-for-software-development-experiments-results-and-future-directions-ac334159ba07?source=rss-bb3486e6f2d------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac334159ba07</guid>
            <dc:creator><![CDATA[Nisaharan Genhatharan]]></dc:creator>
            <pubDate>Fri, 13 Dec 2024 04:41:41 GMT</pubDate>
            <atom:updated>2024-12-13T17:25:59.936Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Abstract</strong><br> The integration of AI agents into software development processes is reshaping how complex engineering tasks are approached. This article explores the experimental advancements in agent-computer interfaces, focusing on state-of-the-art platforms like OpenHands and SWE-agent. These agents leverage large language models to perform tasks such as code generation, debugging, and repository navigation with increasing efficiency. We discuss key experiments demonstrating significant performance improvements, including pass@1 rates of 87.7% on HumanEvalFix and a 6.7-fold increase in resolution rates over retrieval-augmented baselines on SWE-bench Lite. By analyzing the challenges of dataset leakage, error propagation, and safety in execution environments, we highlight how innovations like sandboxing and linting guardrails enhance agent reliability. Finally, we reflect on future directions, such as multi-agent collaboration and expanded interaction paradigms, offering insights into the transformative potential of AI-driven software engineering.</p><p><strong>Introduction</strong></p><p>The advent of AI agents capable of performing software development tasks is revolutionizing the software engineering landscape. With the rise of large language models (LLMs), these agents are becoming proficient in tasks such as code generation, debugging, and repository navigation, often matching or exceeding human efficiency in specific contexts. Platforms like OpenHands and SWE-agent have demonstrated the potential to transform software engineering by enabling autonomous agents to interact with digital environments, execute complex workflows, and produce actionable outcomes.</p><p>OpenHands, an open-source platform, showcases the potential of AI agents to emulate the roles of human developers by integrating capabilities such as code writing, command execution, and web browsing within a securely sandboxed environment. Similarly, SWE-agent leverages custom agent-computer interfaces to facilitate efficient interaction with codebases, yielding significant improvements in solving software engineering benchmarks like SWE-bench and HumanEvalFix. These systems not only demonstrate the viability of agents as digital developers but also underline the need for specialized tools and interfaces tailored to their unique capabilities and constraints.</p><p>This article focuses on the experimental evidence and results supporting these advancements. By examining metrics such as pass@1 rates, benchmark performances, and the impact of innovative interface designs, we aim to highlight the strengths and limitations of current AI agents in software engineering. Furthermore, we address critical challenges, including dataset leakage, error propagation, and execution safety, and explore solutions implemented by platforms like OpenHands and SWE-agent. Finally, we discuss the broader implications and future directions for deploying AI agents in increasingly complex software development environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/522/1*IvkAtUH8cY0Q41kXayz96A.png" /><figcaption>Figure 1: Showcases the design of its Agent-Computer Interface, emphasizing its simplicity and agent-centric design.</figcaption></figure><h4><strong>2. Key Experiments and Results</strong></h4><h4>2.1. Performance of OpenHands on Benchmarks</h4><p>OpenHands, an open-source platform for developing AI agents, has been rigorously tested on various benchmarks to evaluate its capabilities in software engineering tasks. Among its key contributions is its ability to act autonomously within a sandboxed environment, leveraging large language models to execute complex tasks such as debugging, repository exploration, and code generation.</p><p>SWE-Bench Performance</p><p>OpenHands demonstrated its prowess on SWE-Bench, a challenging benchmark designed to test the software engineering capabilities of AI agents. SWE-Bench comprises real-world tasks extracted from GitHub repositories, requiring agents to localize bugs, propose fixes, and verify functionality. By incorporating an event-stream-based architecture and leveraging the CodeAct framework, OpenHands achieved high-resolution rates across diverse test scenarios. These results illustrate its ability to navigate and manipulate complex codebases effectively, often surpassing traditional approaches in both accuracy and efficiency.</p><p>HumanEvalFix Results</p><p>On HumanEvalFix, OpenHands showcased its efficiency in debugging and fixing short code snippets. The platform achieved impressive pass@1 rates by generating high-quality solutions that passed unit tests on the first attempt. This performance demonstrates the ability of OpenHands to excel in localized, short-form programming tasks while maintaining consistency across diverse input cases.</p><p>Comparison with Baselines</p><p>When compared to retrieval-augmented generation (RAG) methods and shell-based interaction baselines, OpenHands consistently outperformed its counterparts. For example, its use of a sandboxed execution environment ensured safe and accurate code evaluations, reducing the risk of cascading errors. Additionally, its modular agent design allowed for seamless integration of specialized tools, further boosting its efficiency and versatility in solving benchmark tasks.</p><p>These findings underscore OpenHands’ ability to act as a reliable and flexible agent for software engineering tasks, providing a robust foundation for further exploration and innovation in AI-driven development environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uo8CMGoayl7rQRVioRPzkQ.png" /><figcaption>Figure 2: Provides an overview of the platform’s architecture, including its event-stream system and runtime environment.</figcaption></figure><h4>2.2. SWE-Agent’s Interface and Performance Metrics</h4><p>SWE-agent represents a significant leap in AI agent design for software engineering tasks by introducing a specialized <strong>Agent-Computer Interface (ACI)</strong>. This interface streamlines interactions between the language model (LM) and the software environment, allowing the agent to efficiently search, navigate, edit, and execute code. SWE-agent’s design prioritizes agent-specific requirements, such as simplified commands, efficient action spaces, and high-quality feedback, which address the limitations of traditional user interfaces designed for humans.</p><p>Interface Highlights</p><p>The SWE-agent interface incorporates several innovative features:</p><ol><li><strong>Custom Commands</strong>: Commands like find_file, search_file, and edit enable efficient localization and modification of code segments. These are tailored to reduce complexity and improve task resolution rates.</li><li><strong>Feedback Mechanisms</strong>: The agent receives concise, action-specific feedback, including context-sensitive updates after each command. This design mitigates error propagation and aids in recovery from mistakes.</li><li><strong>Guardrails</strong>: Integrated syntax checkers prevent invalid edits from being applied, helping the agent maintain the integrity of the codebase.</li><li><strong>File Viewer Integration</strong>: The interface allows agents to view specific sections of files, facilitating targeted edits while conserving computational resources and memory.</li></ol><p>Performance Metrics</p><p>SWE-agent’s performance has been rigorously evaluated on benchmarks like SWE-bench and HumanEvalFix, showcasing its ability to solve complex software engineering tasks.</p><p><strong>SWE-Bench Performance</strong>:</p><ul><li>On the full SWE-bench test set, SWE-agent achieved a resolution rate of <strong>12.47%</strong>, significantly higher than baseline models like retrieval-augmented generation (RAG), which scored only <strong>3.79%</strong> with Claude 3 Opus.</li><li>For SWE-bench Lite, SWE-agent achieved a resolution rate of <strong>18.00%</strong>, a substantial improvement over the Shell-only baseline, which achieved <strong>11.00%</strong>.</li></ul><p><strong>HumanEvalFix Results</strong>:</p><ul><li>SWE-agent achieved a <strong>pass@1 rate of 87.7%</strong> for Python tasks, demonstrating state-of-the-art performance in debugging and short-code generation tasks. This performance far exceeds traditional models, which typically hover below 50%.</li></ul><p><strong>Cost-Efficiency</strong>:</p><ul><li>Despite its high success rates, SWE-agent maintained a competitive average API inference cost of <strong>$1.59 per instance</strong>, making it not only effective but also cost-efficient compared to retrieval-augmented baselines.</li></ul><p>Improvements Over Baselines</p><p>SWE-agent’s interface provides significant advantages over non-interactive models and shell-based systems:</p><ul><li><strong>6.7x Increase in Resolution Rates</strong>: Compared to RAG, SWE-agent consistently resolved more tasks, especially in long-context and multi-step problem-solving scenarios.</li><li><strong>Improved Error Recovery</strong>: Guardrails and structured feedback mechanisms enhanced the agent’s ability to recover from failed edits and avoid cascading errors.</li><li><strong>Scalability</strong>: SWE-agent’s modular design allowed it to integrate seamlessly with larger environments, supporting tasks that span across multiple files and repositories.</li></ul><p>These results highlight SWE-agent’s capacity to redefine how AI agents tackle complex software engineering tasks, providing a practical and effective approach to integrating language models into real-world development workflows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/833/1*-hF7EcaES8Ykoxlhi2C-5Q.png" /><figcaption>Table 1: The agent outperforms baselines on SWE-Bench in terms of % Resolved.</figcaption></figure><h3>3. Comparison and Insights:</h3><h3>3.1 Comparison of Performance Metrics: SWE-Agent vs. Other Frameworks</h3><p>The effectiveness of SWE-agent in software engineering tasks becomes evident when its performance is compared against other prominent frameworks, such as retrieval-augmented generation (RAG) models and Shell-only agents. Key metrics like pass@1 and % Resolved underscore the significant strides SWE-agent has made in the field.</p><h4>Pass@1 Metrics</h4><p>The pass@1 metric reflects the agent’s ability to produce a correct solution on the first attempt, making it a critical indicator of reliability and efficiency.</p><ul><li><strong>HumanEvalFix</strong>: SWE-agent achieved an <strong>87.7% pass@1 rate</strong> for Python tasks, far outstripping RAG and Shell-only baselines, which generally achieve less than 50%. This highlights SWE-agent’s ability to generate high-quality solutions with minimal errors.</li></ul><h4>Resolution Rates on SWE-Bench</h4><p>SWE-agent demonstrated superior performance on the SWE-Bench benchmark, a comprehensive testbed for evaluating software engineering tasks across real-world scenarios.</p><p><strong>SWE-Bench Full Test Set</strong>:</p><ul><li>SWE-agent: <strong>12.47% Resolved</strong></li><li>RAG Baseline: <strong>3.79% Resolved (Claude 3 Opus)</strong></li><li>Shell-only Agent: Not available for the full set but underperformed on subsets.</li></ul><p><strong>SWE-Bench Lite</strong>:</p><ul><li>SWE-agent: <strong>18.00% Resolved</strong></li><li>Shell-only Agent: <strong>11.00% Resolved</strong></li><li>RAG Baseline: <strong>4.33% Resolved (Claude 3 Opus)</strong>.</li></ul><p>Cost-Efficiency</p><p>Despite its advanced capabilities, SWE-agent maintained a competitive average cost per resolved instance:</p><ul><li>SWE-agent: <strong>$1.59</strong> on SWE-Bench Lite.</li><li>RAG models, though less effective, were marginally cheaper, but their lower resolution rates rendered them less efficient overall.</li></ul><h4>Insights from the Comparison</h4><ol><li><strong>Superior Efficiency</strong>: SWE-agent consistently outperformed its competitors by leveraging a specialized Agent-Computer Interface (ACI) tailored for efficient task execution.</li><li><strong>Error Recovery Advantage</strong>: Unlike RAG models, SWE-agent’s integrated guardrails and informative feedback mechanisms allowed it to recover from errors effectively, reducing failure propagation.</li><li><strong>Scalability and Context Handling</strong>: SWE-agent’s ability to manage long-context tasks and multi-file repositories made it particularly effective on the challenging SWE-Bench dataset, which requires a deep understanding of interconnected codebases.</li></ol><p>These comparisons underscore SWE-agent’s dominance in addressing the intricate demands of software engineering tasks, setting a new benchmark for AI agents in this domain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/737/1*PaiMon5iCNPHe24X4CFTiQ.png" /></figure><p>Table 2: compares pass@1 rates, showcasing SWE-agent’s dominance in code generation tasks.</p><p>As shown, SWE-agent’s advanced capabilities enable it to resolve nearly three times as many tasks on the full SWE-Bench dataset compared to RAG baselines and outperform Shell-only agents by a significant margin on SWE-Bench Lite.</p><h4>Performance and Cost Efficiency</h4><ul><li>Average cost per resolved instance on SWE-Bench Lite was $1.59 for SWE-agent, compared to $0.25 for RAG with Claude 3 Opus, but SWE-agent resolved substantially more tasks. This is depicted in cost-performance charts in Section 5 of the SWE-agent paper​.</li></ul><ol><li>Outperforming Baselines: SWE-agent’s performance metrics, such as pass@1 rates and % Resolved, demonstrate its dominance across multiple benchmarks.</li><li>Cost Justification: While marginally more expensive, SWE-agent’s success rates justify the higher costs, particularly in long-context and complex tasks.</li><li>Agent-Computer Interface Impact: Visual insights confirm that SWE-agent’s interface innovations, such as efficient command structures and robust feedback mechanisms, directly contribute to its superior performance.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/282/1*tUo7vuamPKD32bkkyhZaBQ.png" /><figcaption>Figure 3: Highlights the agent’s consistent pass@k performance trends on SWE-Bench Lite</figcaption></figure><h3>4. Challenges and Solutions:</h3><h4>4.1 Challenges in AI Agents for Software Development</h4><p>The implementation of AI agents in software engineering introduces several challenges that must be addressed to ensure reliability, accuracy, and safety. Key challenges include <strong>dataset leakage</strong>, <strong>error propagation</strong>, and <strong>execution safety</strong>, which significantly impact the performance and trustworthiness of these agents.</p><h4>1. Dataset Leakage</h4><p>Dataset leakage occurs when training data inadvertently overlaps with testing or evaluation data, artificially inflating performance metrics. This issue is particularly prevalent in benchmarks like HumanEval, where leaked datasets allow AI models to perform better than they would in novel scenarios.</p><ul><li><strong>Impact</strong>: Leakage undermines the validity of evaluation metrics, making it difficult to assess the true capabilities of AI agents. For example, in HumanEval and ARCADE datasets, agents performed disproportionately well on instances that overlapped with their training data.</li><li><strong>Solution</strong>: OpenHands and SWE-agent address this by employing datasets like SWE-Bench, which are curated to minimize overlap. Tools such as LiveCodeBench evaluate the performance on entirely novel tasks, offering a more realistic measure of agent capabilities.</li></ul><h4>2. Error Propagation</h4><p>AI agents, especially those engaged in multi-step tasks like code editing and debugging, are prone to error propagation. A single mistake in one step can cascade, leading to further errors and reduced task completion rates.</p><ul><li><strong>Impact</strong>: In SWE-Bench, error propagation significantly lowered resolution rates for baseline models. For instance, when agents introduced syntax errors during edits, subsequent actions often failed due to unresolved issues.</li><li><strong>Solution</strong>: SWE-agent mitigates error propagation through integrated <strong>guardrails</strong>, such as syntax checkers and linting tools, which block erroneous edits from being applied. This allows the agent to recover quickly by prompting re-edit attempts. Informative feedback mechanisms further enhance the agent’s ability to correct errors without escalating failures.</li></ul><h4>3. Execution Safety</h4><p>Agents tasked with executing code or interacting with repositories risk causing unintended consequences, such as overwriting critical files, introducing security vulnerabilities, or executing harmful operations.</p><ul><li><strong>Impact</strong>: Safety concerns are particularly acute in autonomous agents, as seen in scenarios where misaligned goals led to unintended system changes. For example, an agent instructed to “make tests pass” might delete the tests instead of debugging the code.</li><li><strong>Solution</strong>: OpenHands incorporates <strong>sandboxing</strong> to create isolated environments where agents can safely execute commands without affecting the host system. SWE-agent further employs credentialing techniques to restrict the agent’s access to sensitive resources, adhering to the principle of least privilege. Additionally, <strong>post-hoc auditing</strong> ensures that agent actions are monitored and logged, allowing developers to identify and rectify any unsafe behavior.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/544/1*Pab_3a2UBqCG5VZcg7z1ZQ.png" /><figcaption>Figure 4: Offers insights into the SWE-Bench dataset, showcasing its real-world relevance and diversity of tasks.</figcaption></figure><p>Addressing challenges such as dataset leakage, error propagation, and execution safety is critical to the success of AI agents in software engineering. Platforms like OpenHands and SWE-agent demonstrate how rigorous dataset curation, error recovery mechanisms, and safety protocols can enhance the reliability and trustworthiness of these systems, paving the way for their broader adoption in real-world applications.</p><h4>4.2 Solutions: Enhancing Reliability and Safety in AI Agents</h4><p>To tackle the challenges of dataset leakage, error propagation, and execution safety, platforms like OpenHands and SWE-agent have implemented robust solutions. These innovations have significantly improved the reliability, accuracy, and safety of AI agents in software engineering tasks.</p><h4>1. Sandboxing in OpenHands</h4><p>One of the standout features of OpenHands is its use of <strong>sandboxing</strong> to ensure safe execution of code and commands.</p><ul><li><strong>What It Does</strong>: Sandboxing isolates the agent’s actions within a secure, containerized environment, such as a Docker container, preventing any unintended interference with the host system.</li><li><strong>How It Works</strong>: The sandboxed runtime allows the agent to execute bash commands, Python scripts, and browser actions without risking harm to external systems. All interactions are mediated through the OpenHands Action Execution API, which ensures that no action extends beyond the predefined environment.</li><li><strong>Impact</strong>: This solution is particularly effective in addressing execution safety concerns, such as accidental overwriting of files or security breaches. It also enables the agent to experiment with potential fixes in a controlled setting, enhancing the trustworthiness of the generated solutions.</li></ul><h4>2. SWE-Agent’s Linting Guardrails</h4><p>SWE-agent incorporates <strong>linting guardrails</strong> to minimize errors during code editing tasks.</p><ul><li><strong>What It Does</strong>: Linting tools automatically validate the syntax and structure of code edits before applying them, preventing the introduction of invalid or buggy changes.</li><li><strong>How It Works</strong>: During an edit operation, the SWE-agent first checks the proposed modification against a set of predefined coding standards. If errors are detected (e.g., syntax violations or logical inconsistencies), the edit is rejected, and the agent is prompted to retry with a corrected action.</li><li><strong>Impact</strong>: This mechanism directly addresses the issue of error propagation by intercepting and resolving mistakes at their source. In benchmarks like SWE-Bench, linting guardrails were shown to significantly improve task resolution rates by reducing cascading failures.</li></ul><h4>3. Credentialing and Post-Hoc Auditing</h4><p>Both platforms employ additional measures to ensure safe and secure operations.</p><ul><li><strong>Credentialing</strong>: By implementing the principle of least privilege, agents are granted only the minimal permissions required to perform their tasks. For instance, SWE-agent uses GitHub access tokens with restricted scopes to prevent unauthorized actions.</li><li><strong>Post-Hoc Auditing</strong>: OpenHands includes logging mechanisms that track all agent actions for later review. This allows developers to audit the agent’s behavior, identify issues, and implement fixes to avoid recurrence.</li></ul><h4>Key Benefits of These Solutions</h4><ol><li><strong>Improved Safety</strong>: Sandboxing and credentialing ensure that agents operate within well-defined boundaries, protecting systems from unintended or malicious actions.</li><li><strong>Error Reduction</strong>: Linting guardrails and feedback mechanisms catch and correct errors early, improving overall task success rates.</li><li><strong>Enhanced Transparency</strong>: Post-hoc auditing provides a clear record of agent actions, fostering accountability and trust in AI-driven systems.</li></ol><p>Innovative solutions like OpenHands’ sandboxing and SWE-agent’s linting guardrails highlight the potential of AI agents to safely and effectively perform complex software engineering tasks. These mechanisms not only address critical challenges but also set a strong foundation for deploying AI agents in real-world environments with confidence.</p><h3>5. Future Directions</h3><p>The evolution of AI agents for software development is paving the way for more sophisticated capabilities and interactions. Building on current advancements, the following areas hold promise for enhancing agent performance, scalability, and safety:</p><h4>1. Multi-Agent Delegation</h4><p>The concept of multi-agent systems opens opportunities for specialized agents to collaborate on complex tasks.</p><ul><li><strong>Opportunities</strong>:</li><li><strong>Task Specialization</strong>: Agents with domain-specific expertise (e.g., debugging, code generation, or web browsing) can work together, with each agent focusing on its strengths.</li><li><strong>Improved Efficiency</strong>: Delegation allows tasks to be divided and conquered, reducing the cognitive load on a single agent and accelerating task completion.</li><li><strong>Current Example</strong>: OpenHands supports multi-agent delegation through its AgentDelegateAction. For instance, a generalist CodeAct agent can delegate web browsing tasks to a specialized BrowsingAgent, improving overall efficiency.</li><li><strong>Future Potential</strong>: Extending this framework to support dynamic collaboration between multiple agents could enable handling of highly complex workflows, such as continuous integration pipelines or multi-step debugging sessions.</li></ul><h4>2. Better Interaction Paradigms</h4><p>Enhancing how agents interact with environments and users is critical for expanding their usability and effectiveness.</p><p><strong>Opportunities</strong>:</p><ul><li><strong>Contextual Awareness</strong>: Agents need to better understand task context and maintain long-term histories to handle projects spanning multiple files and iterations.</li><li><strong>Interactive Interfaces</strong>: Development of richer interaction paradigms, such as mixed-modal interfaces combining text, code, and visual representations, can make agents more intuitive to use.</li><li><strong>Real-Time Feedback</strong>: Introducing more responsive feedback loops during agent execution can allow users to refine or redirect agent actions mid-task.</li><li><strong>Current Example</strong>: SWE-agent’s concise feedback and well-structured command responses are an early step toward such paradigms. However, more intuitive, graphical interfaces akin to human-oriented IDEs remain an area for exploration.</li><li><strong>Future Potential</strong>: Integrating agents with IDE plugins or web-based dashboards could transform them into seamless collaborators for human developers, offering suggestions, explanations, and even learning resources.</li></ul><h4>3. Enhanced Safety Measures</h4><p>Safety remains a paramount concern as AI agents gain more autonomy and capability.</p><p><strong>Opportunities</strong>:</p><ul><li><strong>Preemptive Safety Mechanisms</strong>: Developing predictive models that assess the potential risk of an agent’s actions before execution.</li><li><strong>Behavioral Auditing</strong>: Implementing continuous monitoring of agents’ actions during runtime, with automated rollback or isolation of unsafe operations.</li><li><strong>Ethical Constraints</strong>: Embedding ethical reasoning capabilities to ensure agents act in ways that align with user and organizational values.</li><li><strong>Current Example</strong>: OpenHands’ sandboxing and SWE-agent’s credentialing are robust foundational measures, but they are largely reactive rather than proactive.</li><li><strong>Future Potential</strong>: More sophisticated approaches could include real-time anomaly detection systems or agent behavior prediction models, ensuring a higher level of security and trust.</li></ul><h4>Broader Implications</h4><p>The future of AI agents in software engineering lies in their ability to scale and adapt to increasingly complex environments. Multi-agent delegation can mimic team-based development processes, better interaction paradigms can enhance usability and productivity, and advanced safety measures can build trust and ensure ethical deployment. Together, these advancements could redefine software development, transforming agents into indispensable collaborators in both individual and enterprise-level projects.</p><p>This article draws upon findings from <em>‘OPENHANDS: AN OPEN PLATFORM FOR AI SOFTWARE DEVELOPERS AS GENERALIST AGENTS’</em> and <em>‘SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering,’</em> offering a synthesis of experiments, results, and future directions in the domain of AI agents for software development.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac334159ba07" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How could Airbnb eventually takeoff while other startups of the same idea couldn’t?]]></title>
            <link>https://medium.com/@nisaharan/how-could-airbnb-eventually-takeoff-while-other-startups-of-the-same-idea-couldnt-438dcf5ac2b0?source=rss-bb3486e6f2d------2</link>
            <guid isPermaLink="false">https://medium.com/p/438dcf5ac2b0</guid>
            <category><![CDATA[airbnb]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[thinking]]></category>
            <dc:creator><![CDATA[Nisaharan Genhatharan]]></dc:creator>
            <pubDate>Tue, 12 May 2020 20:10:19 GMT</pubDate>
            <atom:updated>2020-10-13T19:14:18.102Z</atom:updated>
            <content:encoded><![CDATA[<p>Unfortunately, I’m not in a position to utilize the phrase “During this period of time …….” to kick off this writeup. Maybe I’m referring to a timeframe which is the post-pandemic season in which we all are safe. So I’m using this specific timeframe to analyze about “How could Airbnb takeoff while other startups of the same idea couldn’t?”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*isU778hPZ8Krzx2c8Irp7A.jpeg" /><figcaption>Airbed and breakfast which is the extension of airbnb</figcaption></figure><p>Airbnb is “a community built on sharing” began its operation in 2008. Now, millions of hosts and travelers choose to create a free Airbnb account so they can list their space and book unique accommodations anywhere in the world. And Airbnb experience hosts share their passions and interests with both travelers and locals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tuEqJ3glVK1UCNacnbLQLA.jpeg" /></figure><p>Airbnb offer arrangement for homestays, or tourism experiences. The company does not own any of the real estates, nor does it host events; it acts as a broker, receiving commissions from each booking. So the risk starts in the basic idea of the startup itself. Over the internet, You are going to invite a complete stranger to sleep in your home and it&#39;s gonna be huge. Basically, no one in their right minds would invest in a service that would allow a stranger to sleep in their home. Why? Because we all have been taught as kids, Stranger = Danger. So the big struggle “airbnb” faced was to build a gigantic trust between people who never met.</p><p>So for accomplishing the aim of turning fear into fun, Airbnb had to work on 2 things.</p><ol><li>The first thing is, A well-designed reputation system is a key for building trust. Because We prefer people who are like us. The more different somebody is less we trust them. That is the natural social bias. But what interesting is what happens when we add reputation into the mix. In this case with reviews, if u get fewer reviews nothing changes. And If u get more reviews everything changes. “High reputation beats high similarity”. The right design will help to overcome one of these trust issues.</li><li>Building the right amount of trust need the right amount of disclosure<br>This is what happens when a guest first messages a host.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JQeoNcqaI55zk9XhgW3Syg.png" /></figure><p>Using the size of the box to suggest the right length airbnb designs for the right amount of disclosure.</p><p>They didn’t invent anything new, hospitality has been around forever. What differentiated airbnb to eventually takeoff is their core belief of “When trust workout right it can be absolutely magical”.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=438dcf5ac2b0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How meditation changed myself?]]></title>
            <link>https://medium.com/@nisaharan/how-meditation-changed-myself-7e379828d54d?source=rss-bb3486e6f2d------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e379828d54d</guid>
            <category><![CDATA[meditation]]></category>
            <dc:creator><![CDATA[Nisaharan Genhatharan]]></dc:creator>
            <pubDate>Wed, 23 Jan 2019 12:05:29 GMT</pubDate>
            <atom:updated>2020-10-13T19:05:43.190Z</atom:updated>
            <content:encoded><![CDATA[<h3>How meditation changed me?</h3><p>“The following article was written for the Meditation module in the 4th semester and received the highest score of 93 marks.”</p><p>Even though I could excel in my studies in my advanced level education and getting selected to engineering faculty at the University of Moratuwa, during my first 3 semesters in university I really had a bad time. After getting selected to my university what I really thought was to concentrate on my studies only and get a first-class degree in Computer engineering. But after coming to university to lead my life here in Colombo, I really struggled a lot to manage my studies, manage my time and to manage my time with my friends, and to manage my time for personality development. It was the first time that I was away from my home and I am the only one to perform my day to day activities. So there is no one to assist me in washing my clothes, ironing them, polishing shoes, and grooming myself. Actually, we were about 10 people boarded in a one-storied house. But I was the one, who found difficulties in managing my time. So I can be seen stressed most of the time. I was struggling to organize my notes and also to study. If I study well one day, the other day I couldn’t. Before 1 week for the semester exam, I came to realize that I didn’t have a proper plan for this 1st semester, which is the most important part of university life. Because 1st semester is the path deciding semester, which determines our future. After realizing this thing I was stressed more and my study rate was decreased. I had faced the exams with much fear. Even I was not feared for my G.C.E A/L exams as I feared for my 1st-semester exams. Since my mind was not steady and my mind was not confident I ended up doing exams not well. So I couldn’t find a place in the Computer Engineering department and I was selected for Electrical engineering.</p><p>2nd semester came and the same thing repeated I hadn’t scored well. So I found that something is lacking in me. I spent some time to figure it out during my 2nd-semester vacations. I could feel that my mind is not peaceful. It was in search of something. I could feel this was the reason why I couldn’t do work perfectly and also to feel insecure about my future. So I understood that I should do some precautions about this problem. The first thing which came to my mind was doing yoga. But at the time when I realized that my mind needs peace, it was at the end of the 2nd-semester vacations. So I knew that there is no way for going to yoga class. Then I came to the university to begin my 3rd semester. I was seeking a way to bring peace in my mind. So one day I got to know that, there is a yoga class being conducted at university in a new gym. So I asked for the details and got to know that yoga class starts at 6 AM weekly twice. But I couldn’t make that happen. Being boarded in Mt.Lavinia I tried it for 2 weeks and ended up giving up the effort.</p><p>But I tried something new in the 3rd semester. I thought stressing only in one thing to make it better is giving stress to me. So I thought to engage in diverse activities. I joined some clubs and sports. I joined the AIESEC club, Gavel club, and tennis. I thought rather than allocating the whole time for studies, splitting the time perfectly for diverse activities will help me to manage time perfectly thereby doing my tasks perfectly. Even though, I had the problem of getting stressed during exams. I start to study from the very first week and ended up not studying during study leave. So I could know that I’m not doing work perfectly when I’m stressed. So I was in search of a remedy.</p><p>4th semester came and I was given the opportunity to select among western music and meditation. I couldn’t figure out the best way to bring peace in my mind. Because some people say we can find internal peace from music and meditation as well. And after discussing with seniors I ended up selecting meditation.</p><p>My learning experiences and my learning from the very first day of mediation The meditation types and the main outlines which we had studied are</p><ol><li>Metta meditation</li><li>Asubha meditation</li><li>Walking meditation</li><li>Breathing meditation</li><li>Body scan meditation</li><li>Phantom leaf effect</li></ol><p>The important struggle which I had to face in the first class was to sit on the floor for about two hours. It doesn’t mean that I hadn’t sat on the floor for a long time. While we were in school we sat on the floor for a long period of time. But it was about six to seven years ago. So being on the floor for a long period of time after a long time span was a difficult task for not only me but also for most of the others.</p><p>The other important problem was to close my eyes. For me personally, I have been suffering a lot to control my sensors and also In case of studies, the problem I face nowadays is to concentrate my mind in studies for a period of at least half an hour. Sometimes any random thoughts strike my mind while studying or I take the phone to scroll newsfeed. So all of a sudden being in meditation class closing my eyes and sitting on the floor was one of a breakthrough in my life. Because I know I won’t do this meditation without a proper start. So in the first class, all I learned was to close my eyes for a certain period of time and to sit on the floor.</p><p>In the next lecture, the main thing we learned was that our lecturer said, we are controlled by others, the purpose of doing this meditation is to gain control of our mind. Because the mind is our master if we use it in a way so as to bring the best version of ourselves it is the victory.</p><p>The meditation we studied in that lecture is Metta meditation, our lecturer said this meditation will help to reduce the anger, hatred, and to develop love and peace with others.</p><p>May I be well and happy,</p><p>May I be free from anger, fear</p><p>May I be free from hatred, sicknesses</p><p>May I be free from sufferings, grief</p><p>May I take care of myself happily?</p><p>First, we all tell this verse to bring a certain peace in ourselves. And then we use the above verse to all living beings in the directions of north, northeast, east, southeast, west, northwest to spread love. Then we said the above verse to those who love me and after that those who are neutral to me and those who are angry with me.</p><p>In the next lecture, our lecturer told us that “this time there are more than 100 students registered for this meditation module. Usually, I lecture for a limited amount of students. Since this time more students registered for this module, I am going to fail at least 10 students ”. After saying this all of our mindsets were changed. So what really happens is we all are emotionally controlled by words.</p><p>What I really learned from that circumstance was, we will continue to suffer if we have an emotional reaction to everything that is said to us. True power is sitting back and observing things with logic. True power is restraint. If words control us that means everyone else can control us. We should breathe and allow things to pass.</p><p>In the next lecture, our lecturer made us think about “what we hear and get to know about the world is it the reality?”. So I will present my ideology and thoughts about the question down here So the almighty God has gifted us with 6 sensing organs in our human body. They are eye, ear, tongue, skin, brain, and nose. By using these organs we see, hear, taste, feel, think, and smell the environment and use our body parts to fulfill our daily routine. Our sensory organs sense the environment and send the information to the brain and the brain process that information and gives instruction to the particular part which should be moved. We are doing our daily works believing our sensory organs. So here I would like to discuss “What we hear and get to know about the world, is it the reality? ” So for this particular question, I will definitely say that What we hear and get to know about the world is not always real. But I am not saying that our sensory organ (ear) had given us wrong information but it may be due to carelessness for example if there is an announcement in a speaker says that the library is not closed tomorrow but carelessly I hear it as the library is closed tomorrow. So here I get to know about the world wrongly and it is not the reality. Another reason for knowing about world wrongly is similar incident happening and our brain is not being able to differentiate between those two things, for example, consider you are in a beauty parlor having makeup and your eyes are closed and there are two persons have similar voice named Jathu and Gowthaman and Jathu came to a parlor and say Hello to you and you hear it and you say “Hello Gowthaman”. So in this incident, our brain was not being able to differentiate between these two persons and got to know about the world wrongly. Aging also paves a way to get the information wrongly and can make us get to know about the world wrongly. So finally I would like to stress one thing that everything that we hear is not always supposed to be true. But we can make our mind so as to let us calm down and make our mind clean so that it will surely pave us a way to differentiate between things and to avoid carelessness. Then in another lecture what we had seen was a person who is blind but he sees the world In the next lecture we studied asubha meditation. Our lecturer said that this meditation will help to reduce the attachment of us from the body. Because when we begin to reduce the attachment of us to the body, our desires will reduce.</p><p>The other important type of meditation we had studied was walking meditation. We inhale air when we lift the feet and exhale air out while landing the feet and walked some distance and then stood there for a certain time period and then observed the breath for some time and then slowly turned back and did the breathing practice while walking. Walking meditation is to overcome the laziness of mind. I often encounter the laziness of mind in the mornings after waking up. So I started doing this meditation in the mornings. I believe by concentrating on walking we can get rid of the laziness of our mind.</p><p>So after that lecture what I try is to bring the walking meditation into my life. So I followed the following steps to bring walking meditation into my life.</p><p>The first step I preferred was to find a location. So I found a lane that allows me to walk back and forth for about eight to ten feet. Because that place should be relatively peaceful, where I won’t be disturbed or even observed (since a slow, formal walking meditation can look strange to people who are unfamiliar with it). I thought to practice walking meditation either indoors or outside in nature. The lane doesn’t have to be very long since the goal is not to reach a specific destination, just to practice a very intentional form of walking where I’m mostly retracing with my steps.</p><p>The next step was to start my steps. I walked about eight to ten steps along the lane I’ve chosen, and then pause and breathe for a certain time period. Then after I’m getting ready, I turn and walk back in the opposite direction to the other end of the lane, where I can pause and breathe again. Then, when I’m ready, I turn once more and continue with the walk. What I have done when my mind wanders is really important here. Because no matter how much we try to fix our attention on any of these sensations, our mind will inevitably wander. But I believe it’s natural. When I try to notice my mind wandering, I simply try again to focus it on one of those sensations. The breathing meditation is a life-changing one for me personally. In this meditation, we had chosen an area of mindfulness. It may be any part of the body. Our mind will create a mental image representing the area of mindfulness. Personally, I believe breathing meditation gives an immediate response to a person in stress. So for my case, I often get stressed when I think about the “to do” list. It often happens to me when I think about it. So in that case I began to bring breathing meditation into my real life. I believe breathing meditation will not only make our self relieve ourselves from stress but also to be in patience.</p><p>The other important type of meditation we practiced is body scan meditation. Starting with our head, bringing our awareness to our body and notice any tension we’re feeling as we practice our body scan meditation. First, we scan our body and identity are there any feeling of tightness in our shoulders, back, neck, or anywhere else then we think about sensations of pain, either sharp, dull, or subtle then we check if there is a feeling of concentrated ‘energy’ around a certain area? After having seated for a minute and noticing what we’re feeling in our body. We may want to start at the top of our head and systematically focus on each area of our body on our way down — scalp, ears, cheeks, neck, shoulders, chest and really noticing what sensations we have in each area. I believe this becomes easier and more automatic with practice to the point where we’ll be able to do this with very little effort. Through the practice, I could feel that body scan meditation is a practice that can be performed daily or even several times a day and can help us learn to identify what we are feeling and where we’re feeling it, and learn to release the stress in our body and mind. Our lecturer introduced a different ideology in the next lecture, he said us to close our eyes while sitting in a posture and imagine how we are sitting on the floor. So how we could be able to imagine our posture? It’s due to our past experience. So when our eyes are closed, it’s not sure, we are perceiving reality. Because when our eyes are closed we are using our experience to create an image of our present conditions. And the imagination process is subjective. It depends on person to person. So eye doesn’t deceive us it’s the brain that deceives us. So what we could understand from this is wisdom is the key to unlock the reality.</p><p>So this is the flow chart shows what happens when we close our eyes and imagine any object. What I believe is bringing walking meditation into my daily life. Now what I could do is slow formal walking, because it tastes good to me. But the thing I would really love to do is bringing walking meditation at any speeds even while running. The reason is, I can fulfill walking meditation through fulfilling my day to day activities</p><p>From childhood in the case when I am not scoring well in exams my teachers would say one thing that, “everybody is born with the same brain but why are you not scoring well?”. But I couldn’t agree with the statement that everybody is born with the same brain. Because I study hard for the exams, as best as I can but I couldn’t score above a certain level, but some students score well in exams. If everyone is born with the same brain, how some persons are being intelligent to score more marks? Throughput these meditation lectures I got to know the answer, everybody is born with the same brain but we have different mentalities. We all have brains to think, but there should be a peaceful environment in our minds that pave way for efficient thoughts. To create a peaceful mind we should find a way to create inner peace. So meditation will help to create inner peace thereby create innovative and efficient thoughts.</p><p>In the learning experience, the first thing which comes to my mind, if anybody asks me a question what have you learned in your meditation class is, “to suppress bad feelings like anger, fear, and doubt and thereby gaining wisdom through controlling our mind” My view about energy changed totally through meditation. There are a lot of aspects of energy, but what I’m going to talk about are human and energy. Think about how we exist in the world. It’s by inhaling oxygen and burning it, and drinking water for fluid flow and eating food. The common thing happening over here is the production of energy. Yes, we are producing energy for ourselves. For what ….. We are producing energy for our day to day lives. It may be for running, sleeping, talking, studying, and thinking. But I believe, we want energy for some other things too. Do some other things mean? Yeah, think of an instance someone is making us angry and we are burning as hell as a hell, quarreling and fighting with him. So, it means we don’t have enough energy over here. If we had energy in that particular incident we will act in a very controlled manner, we will notice everything around us, how people speak, the choice of words they use. But through meditation, I understood that enough energy won’t come at all very quickly and easily. It needs practice and experience. For this to happen, our minds should be very clean. It means there should not be anything running in our minds. It may be our position in our society, our problem with the neighbor aunty who has hidden the tennis ball which hit her home glass, or with an ex-lover, or with any random traitors, not running in our mind. So it is not all about controlling and staying calm. It’s about how we utilize our energy to use our minds clean and use our intelligence. So for this purpose, we can use breathing practice, I believe. We all produce energy by inhaling air. It means our breathing rate influences our energy production. So if we have control of our breathing, it means we have proper control of our energy flow in our body. What I believe is, by doing breathing practice we can gain control of the energy flow in our body which will pave way for peace in our mind.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e379828d54d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[It’s all about how we resist]]></title>
            <link>https://medium.com/@nisaharan/its-all-about-how-we-resist-e500f9f9b398?source=rss-bb3486e6f2d------2</link>
            <guid isPermaLink="false">https://medium.com/p/e500f9f9b398</guid>
            <dc:creator><![CDATA[Nisaharan Genhatharan]]></dc:creator>
            <pubDate>Fri, 28 Dec 2018 01:32:15 GMT</pubDate>
            <atom:updated>2020-10-13T19:12:47.718Z</atom:updated>
            <content:encoded><![CDATA[<p>One of the events which affected me badly in recent times is the sexual harassment of a woman in front of her husband in Jaffna. It was such a brutal one. It will be so difficult for that family to come out of that accident. Since the Nithya&#39;s case, I ask myself what is a typical boy in Jaffna I can do to make a full stop to these types of brutalities. I prefer that each and every person’s data be stored as a document with the government. To be speaking more frankly the government should implement a system which is similar to the system which is followed in the USA. They follow a system called SSN which is an abbreviation for Social Security Number. It’s similar to our NIC. But it contains some more data. It contains a person’s fingerprint, no of times he went to jail, his working place, his job, and a lot more. But they don’t have a DNA sample of each. This data is only accessible by the government. So if we introduce a system that is similar to SSN, including DNA sample of a person. In the instances of rape scenes, the police can identify the criminal’s DNA and with the help of the government they can catch up the criminal. I’m sure in a country like Sri Lanka where these types of methods are possible only when the government acts linearly and without ignorance. And I’m again pretty sure it will take a long span of time. But we should implement these kinda methods to protect our sisters in Jaffna.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e500f9f9b398" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[So why we are using our energy for…..]]></title>
            <link>https://medium.com/@nisaharan/so-why-we-are-using-our-energy-for-f03423c61a32?source=rss-bb3486e6f2d------2</link>
            <guid isPermaLink="false">https://medium.com/p/f03423c61a32</guid>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Nisaharan Genhatharan]]></dc:creator>
            <pubDate>Fri, 17 Aug 2018 19:12:53 GMT</pubDate>
            <atom:updated>2020-10-13T19:10:12.032Z</atom:updated>
            <content:encoded><![CDATA[<p>So being an electrical engineering student I’m not going to talk about automation, instrumentation, installations, or electric motor used in Tesla Model S. But something about energy. But wait, I could hear your voices saying “since you are studying power electronics and power systems and sustainable energy you may have ideas about energy and power”. But that’s not the key factor for me to write this article over here. It may be my failures, weakness, my inabilities, or disabilities to adapt to a new environment or a new situation or hard people. You may ask whom you are referring to as hard people. It may be your new lecturer standing very strict in your lecture hall without letting you sleep and making a very sure attendance or it may be the very next guy sitting right next to you whom you don’t like to talk.</p><p>So let’s stick to the theme. So we are talking about energy. There are a lot of aspects of energy, but what we are going to talk about are human and energy. Think about how we exist in the world. It’s by inhaling oxygen and burning it, and drinking water for fluid flow and eating food. The common thing happening over here is the production of energy. Yes, we are producing energy for ourselves. For what ….. We are producing energy for our day to day lives. It may be for running, sleeping, talking, studying, and yeah some boys are saying for masturbation too. So if anyone comes and asks you about how energy is dissipated from our body we will say these factors. But do you think energy is wasted only for physical activities? Yeah, you are saying even for thinking Nah? But I believe, we want energy for some other things too. Do some other things mean?</p><p>Yeah, think of an instance someone is making your angry and you are burning as hell as a hell, quarreling and fighting with him. So, it means you don’t have enough energy over here. If you had energy in that particular incident you will act in a very controlled manner, you will notice everything around you, how people speak, the choice of words they are using. But that enough energy won’t come at all very quickly and easily. It needs practice and experience. For this to happen, your mind should be very clean. It means there should not be anything running in your mind. It may be your position in your society, your problem with your neighbor aunty who has hidden your tennis ball which hit her home glass, or with your ex-lover, or with any random traitors, not running in your mind.</p><p>So it is not all about controlling and staying calm. It’s about how we utilize our energy to use our minds clean and use our intelligence.</p><p>So it’s just an experimental random first essay, hope I will do more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f03423c61a32" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>