<?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 Sandeep Mahendru on Medium]]></title>
        <description><![CDATA[Stories by Sandeep Mahendru on Medium]]></description>
        <link>https://medium.com/@mahendru.sandeep?source=rss-b092604079f6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*T8HHt6FbOGTwDUYmZAWJtA.jpeg</url>
            <title>Stories by Sandeep Mahendru on Medium</title>
            <link>https://medium.com/@mahendru.sandeep?source=rss-b092604079f6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 23:55:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mahendru.sandeep/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[Proposal for the Global Backbone of the Agentic Web]]></title>
            <link>https://medium.com/@mahendru.sandeep/proposal-for-the-global-backbone-of-the-agentic-web-718339ae25f2?source=rss-b092604079f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/718339ae25f2</guid>
            <category><![CDATA[registry]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sandeep Mahendru]]></dc:creator>
            <pubDate>Fri, 03 Apr 2026 03:07:37 GMT</pubDate>
            <atom:updated>2026-04-03T03:07:37.752Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><h3>Setting the Stage for the Agentic Web</h3><p>We have reached the “Wild West” era of artificial intelligence. While the world is mesmerized by the latest LLM benchmarks, a much more profound shift is happening beneath the surface: the transition from writing explicit code to orchestrating autonomous intent. However, we have a massive infrastructure problem. We are currently building millions of specialized agents, yet we have no global address book to find them, no shared language to verify their skills, and no scalable security layer to ensure they aren’t actively sabotaging our workflows. We are living through the <strong>“Cisco Moment”</strong> of the Agentic Web — a pivotal era where the winners won’t be defined by who builds the smartest model, but by who builds the universal backbone that allows trillions of models to discover, trust, and execute alongside one another <strong>without introducing catastrophic latency</strong>.</p><h3>The Agentic Web at a Glance</h3><ul><li><strong>The Paradigm Shift:</strong> We are moving from <strong>“Code-First”</strong> development to <strong>“Intent-First”</strong> orchestration. The future is a global “quilt” of specialized agents, not a single monolithic model.</li><li><strong>The Discovery Brain:</strong> Traditional search is fading away. The new “Brain” uses <strong>Semantic Intent Parsing</strong> to decompose human requests into multi-agent workflows in real-time. Instead of a centralized bottleneck, a standardized metadata routing protocol allows secure, local orchestrators to decompose human requests into multi-agent workflows while maintaining strict context isolation.</li><li><strong>Report Cards, Not Business Cards:</strong> The industry is pivoting from self-declared <strong>A2A Agent Cards</strong> to cryptographically verified <strong>AgentFacts</strong>. Trust is no longer a “claim” or a static audit; it’s an audited metric signed by third-party authorities and constantly tested via automated prompting and local consensus.</li><li><strong>The Security Mandate:</strong> In a federated web, <strong><em>“Trusted”</em></strong> agents can still lie, hallucinate, or be sabotaged. <strong>Byzantine Fault Tolerance (BFT)</strong> is the mandatory “Anti-Sabotage” layer that uses triangulation to ensure system integrity. Because traditional consensus mechanisms are too slow for real-time inference, economic staking and slashing provide the mandatory “Anti-Sabotage” layer to dis-incentivize bad actors.</li><li><strong>Cognitive Liquidity:</strong> We are entering the era of <strong>“AI-NPM” and intelligence subscriptions.</strong> AI Firms now need to provide the enterprise gateway where developers securely stream or “import” global intelligence libraries, verified agent capabilities into low-code environments and associated tooling and dashboards.</li></ul><h3>Transition from Agent development to Describing Intent</h3><p>We are witnessing a fundamental paradigm shift from traditional “code-first” development to a visionary era of “intent-first” orchestration. Instead of building monolithic software silos, developers are now focused on defining high-level goals that are executed by a “quilt” of autonomous, specialized agents. This transition moves the cognitive burden from the human developer to an underlying intelligent infrastructure, signaling the arrival of the “Cisco Moment” for the global Agentic Web.</p><p><strong>The Evolution of Agentic Development</strong></p><p><em>The industry is rapidly shifting from “writing code” to “describing tasks.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s67CZWUYB-eZ-mxw7dmt-g.jpeg" /></figure><h3>The Problem: The Discovery and Trust Gap</h3><p>Despite the rapid proliferation of specialized AI models, the ecosystem currently lacks a unified global address book to facilitate efficient discovery. Without a standardized naming service, <strong><em>complex user intents</em></strong> remain trapped within proprietary silos rather than being routed to the<strong><em> best-in-class experts</em></strong>. Furthermore, the absence of a verifiable trust layer makes it impossible for orchestrators to distinguish between honest, high-performing agents and malicious or incompetent actors.</p><p><strong>The Missing infrastructure</strong></p><p>Today both the ecosystem of agent collectives and the infrastructure (Federated registries) do not exist. The ecosystem is getting richer by the day and very soon, we would have agents, mcp and tools for <strong><em>almost every automation need.</em></strong></p><p><em>By then, the AI firm that provides the infrastructure, to enable the discovery of the best in class agents would be able to monetize.</em></p><p>In short <strong>infrastructure</strong> = (Federated registry construct + Brain (mechanism to translate the intent into a semantic search across the federated registries))</p><p><strong>The Need for the “Brain” &amp; Global Discovery</strong></p><p><em>Complex cognitive tasks require an orchestrator that can navigate the “Agentic Web.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TwffE6YasQAQ6k9_wZggzg.jpeg" /></figure><h3>The Architecture: A2A, NANDA, and the “Quilt”</h3><p>To bridge this gap, a layered architectural framework utilizing the A2A protocol and the NANDA “Quilt” index is required to facilitate decentralized interoperability. While A2A provides the essential communication standard or “business card” for agents, NANDA functions as a federated switchboard that indexes these capabilities across a <strong><em>decentralized set of federated registries.</em></strong> This model ensures that agents remain independently governed while being discoverable through a lightweight metadata layer, prioritizing an open ecosystem over centralized platform control.</p><p><strong>A2A as the standard for Agent communication</strong></p><p>The A2A protocol provides a secure, flexible, and open standard for agent interoperability through three core pillars:</p><ul><li><strong>Web Standards Foundation:</strong> Rather than inventing bespoke communication methods, A2A is built on established internet standards like JSON-RPC 2.0 over HTTP/S and Server-Sent Events (SSE) for streaming.</li><li><strong>Enterprise-Grade Security:</strong> It treats agents like standard applications, securing them with widely adopted authentication schemes such as OAuth 2.0, mTLS, and API keys. It strictly scopes data access to the caller’s authorization boundaries and even supports dynamic, mid-task authentication if an agent needs new credentials during execution.</li><li><strong>Modular Extensibility:</strong> To remain future-proof, A2A features a formal extension system. Developers can declare custom or vendor-specific capabilities in an Agent Card, which clients can seamlessly opt into using standard HTTP headers like A2A-Extensions without altering the core protocol</li></ul><p><em>A2A provides the standardized language (Agent Cards). In order to perform the complex semantic discovery as described by the scenario outlined in the schematic, the AI firms would need to combine A2A with an intelligent Agent Naming Service and a federated architecture like NANDA or AGNTCY.</em></p><p><strong>Capability-Based Discovery &amp; Federated Registries</strong></p><p><em>Bridging the gap between standard communication and intelligent discovery.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vKNBhFKhvpBDtuxnl8OoJQ.jpeg" /></figure><h3>The Intelligence Layer: The “Brain” and Semantic Search</h3><p>At the heart of this ecosystem lies the “Brain” . The Brain functions as a sophisticated, LLM-powered intent parser. Unlike traditional keyword-based search engines, the Brain utilizes semantic decomposition to translate natural language requests into structured, multi-agent workflows. By analyzing real-time context and operational constraints, the brain identifies the precise combination of experts — from logistics to finance — needed to execute a complex task. The schematic below illustrates a high level constitution of the brain. It defines the blueprint. An AI enterprise would need to work on a detailed task decomposition, security components (reputation and certification) and an ontology.</p><p><strong>“Brain”: Intent &amp; Global Discovery</strong></p><p><em>Establishing an intelligent service for best-in-class agent procurement.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4rFwFFmx_RUFC7BlJMEaAA.jpeg" /></figure><h3>The Economic Future: Cognitive Liquidity and AI-NPM</h3><p>The long-term economic value of this infrastructure lies in the creation of “Cognitive Liquidity” through a global, reusable AI-NPM library model. By integrating discovered agents directly into low-code environments like Gemini Enterprise, developers can <strong>“import”</strong> <em>verified intelligence</em> as easily as they import software libraries today.</p><p><em>Cognitive Liquidity — importing intelligence as easily as an npm library</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GLnurmcPLx972Ex91xfSKw.jpeg" /></figure><h3>The Security Mandate</h3><p>In an open federated web, identity is not a proxy for integrity, necessitating a robust security model based on Byzantine Fault Tolerance (BFT). We must move beyond self-declared Agent Cards toward cryptographically signed<strong><em> “AgentFacts”</em></strong> that serve as a verifiable, audited report card for an agent’s performance. Agent facts would serve as an extension to A2A and provide a reputation summary for the agent usage history. This report will be certified by an authority using the W3C verifiable credentials standard. [https://www.w3.org/TR/vc-data-model-2.0/]</p><p>An added security layer based on the BFT approach is required on top of the certified AgentFacts. By triangulating data through quorum voting and independent gateway telemetry, the system can detect “Byzantine” lies and sabotage even from supposedly trusted or established actors.</p><p><em>“A Secure Federated Registry is like a ‘No Fly List’ — it keeps known bad actors out. But </em><strong><em>Byzantine Fault Tolerance</em></strong><em> is like ‘In-Flight Security’ — it protects the passengers even if someone with a valid ticket turns out to be a threat.</em></p><p><em>From basic identity trust to cryptographically audited behavior</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*WRcGl-rrsFxzbwnRD8GenQ.png" /></figure><p><strong>Agent Facts</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P-XPKL9bkvD01ixDntvxfA.jpeg" /></figure><h3><strong>The Open Standard Mandate</strong></h3><p>Ultimately, the success of the Agentic Web depends on the adoption of open standards to prevent vendor lock-in and encourage large-scale ecosystem participation. Governance through neutral bodies like A2A and Linux Foundation ensures that the <strong><em>“Global Address Book”</em></strong> remains a public good rather than a proprietary, closed-loop silo. By leading this initiative, an AI Company establishes themselves as a standard router and trust anchor for the next generation of the internet’s autonomous economy.</p><h3>Closing : The “So What?”</h3><p><em>“Trust” mandate and the power of open standards.</em></p><p>The future of AI is not a single, monolithic “God Model” that answers every question; it is a global, federated quilt of trillions of specialized agents collaborating in real-time. As we move toward this <strong><em>“Agentic Web,</em></strong>” the traditional metrics of cloud success are being rewritten. It is no longer enough to provide raw compute; the next decade belongs to the platform that provides <strong>Cognitive Liquidity</strong> — the ability to “import” verified intelligence as easily as we import a software library today. By championing open standards like A2A and NANDA, we are laying the plumbing for the next era of global commerce. In this new economy, <strong><em>discovery, collective intelligence and verifiable trust are the ultimate frontier.</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qeE7pUUj6XC9uT5VGo93LA.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=718339ae25f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Distributed systems and Cloud native approaches]]></title>
            <link>https://medium.com/@mahendru.sandeep/distributed-systems-and-cloud-native-approaches-365b397aab1d?source=rss-b092604079f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/365b397aab1d</guid>
            <category><![CDATA[distributed-systems]]></category>
            <category><![CDATA[cloud-native]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[Sandeep Mahendru]]></dc:creator>
            <pubDate>Tue, 25 Dec 2018 20:02:39 GMT</pubDate>
            <atom:updated>2018-12-25T20:02:39.566Z</atom:updated>
            <content:encoded><![CDATA[<p>I have worked on designing and building distributed systems for quite some time. More recently, organizations want to build these systems on public clouds or have a plan for hybrid cloud. The reasons and motivations are well known.</p><p>However there are various solution approaches to realize the above. Also within each approach, there are various frameworks, libraries, tooling and public cloud offerings.</p><p>This article discusses various <strong><em>design motivations</em></strong> and <strong><em>associated options</em></strong><em> </em>for accomplishing those.This space is evolving at a very rapid pace and a new concepts and frameworks are being created.There are good chances that new techniques and products would appear in future, that would not be covered in this article. The focus of this article is not to provide in-depth understanding of each discussed framework or concept. Instead it talks about their <strong><em>respective role</em></strong> in building an ecosystem of distributed services using <strong><em>cloud native principles.</em></strong></p><p><strong>Characteristics of cloud native applications</strong></p><p>The definition of Cloud Native is very subjective. Based on my reading and experience, cloud native approach emphasizes certain attributes for building and deploying applications on an infrastructure.Cloud native applications should have the following characteristics:</p><p><em>Micro services based design<br> Provide service health and instrumentation<br> Declarative<br> Automated Application communication</em></p><p>The above traits enable the system to be more agile, flexible, improved maintainability and better <em>portability</em> across different cloud platforms.</p><p><strong>Micro Services design objectives</strong></p><p>Micro service based design is a topic in it’s own right and is not the focus of this article. But, it is important to understand the major design goals:</p><p>a] Distributed systems with well defined boundaries . <br>b] Each subsystem manages the state of an important business entity<br>c] Subsystems do not share the state and communicate using events<br>d] Have very good support for instrumentation at a system level and it each subsystem.<br>e]Support for automated deployments and system integration.</p><p>A few <em>key benefits</em> of the above approach are:</p><p>a] Each subsystem can be enhanced and deployed independently,avoiding complete system shutdown.<br>b] Each subsystem can scale independently.<br>c] Each subsystem can be implemented in different language based on the required use-case.<br>d] Failures can be localized to the subsystem boundary and avoid cascading failures. Design mechanisms like circuit breaker and promises can be used to build this semantic.</p><p><strong>Challenges with Micro Services</strong></p><p>The challenges with a micro services based systems are primarily :</p><p>a] Cluster management. There are various design goals associated with this aspect. The details have been described under the section: <em>Cluster management</em></p><p>b] Building a good monitoring and instrumentation capabilities for the complete system and provide detailed metrics view for each subsystem. The capabilities required have been explained under the <em>telemetry section.</em></p><p><strong>Cluster management</strong></p><p>In my experience, we had to <strong><em>engineer</em></strong> and build all of the above outlined from <strong><em>ground up</em></strong> in an organization’s data center. If the management had budget, this would be on a dedicated infrastructure or otherwise it would run on a shared infrastructure. Generally, the latter scenario is more prevalent.</p><p>Today, The focus is towards an <strong><em>automated approach</em> </strong>towards managing cluster of services, instead of building these from ground up. Secondly, the approach should also provide <strong><em>policy based [declarative] mechanisms</em></strong> to build and apply changes to the cluster.</p><p>Today, most of the <strong><em>cluster engineering capabilities</em></strong> have been automated and managed by Kubernetes. Some of these capabilities are: <br><em>Management</em> <em>of micro services cluster as defined by the development team</em><br> a] Docker compose or Minikube for building local simulated dev sandbox for each developer. <br> b] Automate the boot strapping of the entire cluster on the required nodes.<br> c] Automates the execution of container based micro services based on configured topology.<br>d ]Policy based Auto scaling<br> e] Rollover updates with zero downtime<br><em>Resiliency</em><br> Maintain specified cluster state based on configured replica sets.<br> <em>Multi Tenant cluster setup</em> via labels and namespaces.<br> <em>Advanced scheduling</em> : Node and Pod Affinity</p><p><strong>Service health</strong></p><p>Basic idea is that the application provides information<strong><em> about it’s state and important metrics.</em></strong> Determining the health , by monitoring the hardware resources is incomplete. In early days, Java applications would expose important application health metrics as JMX beans. For example app server frameworks exposed application lifecycle states like active, suspended, ready , busy . Compute frameworks exposed details about the No of Requests currently being serviced, No of Requests waiting, Size of the Queue buffer. However, <strong><em>most of this required customer coding.</em></strong></p><p><strong><em>Instrumentation and metrics are </em></strong>important information about the system that is necessary for the development team to make future design choices and enhancement to the system. These could be important benchmark parameters for the complete system and for each subsystem like No of requests the system can process per unit of time [<em>Throughput]</em>, <em>Error Rat</em>e [No of transaction in error state], How long does it to process a transaction and respond.[<em>Application latency</em>]</p><p>There are many frameworks and instrumentation libraries that can be used to enable the above, instead of coding these from Ground up.</p><ul><li>Zookeeper provides service metadata and more recently Kubernetes’ etcd http endpoint provide each service’s health status.</li><li>Spring Boot provides out of the box JMX information using Spring Actuator</li><li>DropWizard metrics library</li><li>Spring Boot’s Micrometer Metrics library API for managing application-specific meters</li><li>AOP based techniques to define required point cuts,backed by an audit cache. Today, frameworks like Spring Cloud Sleuth provide this out of box.</li><li>Istio’s distributed Flow Tracing provides metrics on each subsystem</li><li><em>Grafana</em> and <em>Prometheus </em>can be configured to work with a micro service based systems to receive the above metrics and display these using charts on a single dashboard.</li></ul><p><strong>Service Mesh</strong></p><p>The primary motivation for service mesh is to <strong><em>automate application specific networking functions</em></strong> or layer 5 communication.The design goal here is to provide mechanisms for defining<em> communication policies</em> between <em>application components.</em></p><p><a href="https://istio.io/"><strong><em>Istio</em></strong></a><strong><em> </em></strong>is a popular choice to build these capabilities. Istio is <strong><em>application routing focused </em></strong>as compared to Kubernetes based routing which operates at network namespace level.</p><p>The important capabilities that Istio provides are:</p><p>a] Policy based routing : Dynamic route configuration, A/B tests and Canaries<br>b] Advanced Load Balancing:<br> Declarative way of specifying timeouts, retries, health checks and circuit breakers. <br>c] Observability: Distributed Flow Tracing.</p><p>Based on the above features, Istio can be leveraged to build:<br> a] Separate deployments in prod and seamless transitions<br> b] Istio can be used to provide policy based integrations between the multiple clusters.<br> c] Istio provides creating custom load balancers that can be configured using policy based routing. The policy definitions are not tied to specific Cloud provider.</p><p><strong>Declarative</strong></p><ul><li>Topology of a distributed system needs to be specified in a declarative manner. Service setup , bootstrap and state definitions should be configured. Kubernetes is designed to address this goal.</li><li>This concept is further extended to cloud environments. A framework needs to interpret an on premise K8 style topology and then realize it on different public clouds. Kops and Terraform together can be used to configure and build a K8 cluster on a cloud.</li><li>Cloud Formation can be used to create a EKS cluster on AWS. But this approach would be very specific to AWS.</li></ul><p><strong>Portability</strong></p><p>Micro services based design and run time management by Kubernetes leads to a more portable design. This is because, the service life cycle , service communication , telemetry and management is done by cloud agnostic frameworks. For example:</p><p><strong><em>Distributed System A</em></strong> is built using AWS specific run time capabilities. Each service specific business function is coded and deployed as a separate Lambda function.The external client calls are intercepted by AWS ELB and Gateway specific deployment. Service life cycle and communication is managed by AWS runtime. Auto scaling configuration are also AWS specific. If , for some reason there is a requirement to move this complete setup or a part of it to another cloud platform, it would require significant effort to re code and redeploy on a different platform.</p><p><strong><em>Distributed System B</em></strong> is built using open source alternatives. Each service is built using Spring Boot for JVM or NodeJs. API Gateway is implemented using frameworks like Spring Cloud or GraphQL and inter service communications leverage Spring boot REST or GRPC. Each service is packaged as a Docker Container and deployed on a K8 platform. Custom App routing is implemented as Istio based policies. Such a system can be ported with relative ease to different public clouds. For example system B could be ported to AWS using EKS or to GCP using GKE.</p><p>System B exhibits the Cloud native characteristics and in essence it create a platform on top of IaaS offerings from public clouds.</p><p>Building a micro-service system from ground up using Kubernetes and Istio can be a time consuming undertaking. It also requires a skilled team.</p><p>A much higher level of abstraction or a Paas based approach to building K8 based cloud native systems is using <a href="https://www.openshift.com/">Openshift </a>or <a href="https://pivotal.io/platform/pivotal-container-service">PCS </a>[Pivotal container service]</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=365b397aab1d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>