<?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 Aarchi Bhanushali on Medium]]></title>
        <description><![CDATA[Stories by Aarchi Bhanushali on Medium]]></description>
        <link>https://medium.com/@aarchi.bhanushai24?source=rss-f79593c83d86------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*KRw4x5fRosWLbqWC</url>
            <title>Stories by Aarchi Bhanushali on Medium</title>
            <link>https://medium.com/@aarchi.bhanushai24?source=rss-f79593c83d86------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 15:48:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aarchi.bhanushai24/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[If AI Could Think Like You: How Intelligent Agents Plan and Make Decisions]]></title>
            <link>https://medium.com/@aarchi.bhanushai24/if-ai-could-think-like-you-how-intelligent-agents-plan-and-make-decisions-b79af667940f?source=rss-f79593c83d86------2</link>
            <guid isPermaLink="false">https://medium.com/p/b79af667940f</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[search-algorithm]]></category>
            <category><![CDATA[intelligent-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Aarchi Bhanushali]]></dc:creator>
            <pubDate>Fri, 07 Aug 2026 05:04:16 GMT</pubDate>
            <atom:updated>2026-08-07T09:37:21.889Z</atom:updated>
            <content:encoded><![CDATA[<p>It’s 8:10 AM. Your lecture starts at 9. You glance outside — overcast, might rain. You mentally run through your options: the auto is faster but pricier, the bus is cheap but might be late if it rains, the train skips traffic entirely but means a longer walk on both ends. You weigh all of it in about four seconds and walk out the door with a decision already made.</p><p>You didn’t just react to being late. You <em>planned</em>. You looked ahead, imagined a few different versions of the next 50 minutes, and picked the one most likely to end well.</p><p>Strip away the human intuition, and that’s precisely the problem AI researchers have spent decades trying to solve: how do you get a machine to do the same thing — look ahead, weigh options, and choose — without a brain, without instinct, and without ever having taken that bus before?</p><h3>The Agent: AI’s Basic Unit of “Thinking”</h3><p>In AI, anything that perceives its environment and acts on it is called an <strong>agent</strong>. That’s a deliberately broad definition — a thermostat is an agent. So is a chess engine. So is the self-driving car navigating actual traffic. What separates a “dumb” agent from an “intelligent” one isn’t the hardware, it’s how much reasoning happens between perceiving and acting.</p><p>To design any agent, AI engineers use a simple framework called <strong>PEAS</strong>:</p><p><strong>Performance measure</strong> defines what “doing well” even means — for a delivery robot, that might be speed, battery efficiency, and never bumping into someone. <strong>Environment</strong> is the world the agent operates in — a warehouse floor, a road network, a game board. <strong>Actuators</strong> are how it acts on that world — wheels, a robotic arm, a steering system. <strong>Sensors</strong> are how it perceives that world — cameras, GPS, pressure sensors, the pixels on a chessboard.</p><p>Once you can answer those four questions for a system, you’ve essentially defined the problem it needs to solve. Everything else — planning, learning, deciding — is built on top of that foundation.</p><h3>From Reflexes to Reasoning: A Spectrum of Agents</h3><p>Not every agent “thinks” the same amount, and this is where things get genuinely interesting.</p><p>At the simplest end sit <strong>reflex agents</strong> — machines that map a current perception directly to an action, with zero memory of the past. A thermostat sensing “too cold” and switching on the heater is a reflex agent. It works, but it’s brittle; it can’t handle anything it wasn’t explicitly told to react to.</p><p>One step up are <strong>model-based agents</strong>, which maintain an internal model of the world so they can reason about things they can’t currently sense. A robot vacuum that remembers which corners of the room it already cleaned — even after temporarily losing track of its exact position — is using a model, not just a reflex.</p><p>Then there’s the real shift: <strong>goal-based agents</strong>. These don’t just react to the present, they simulate possible futures and pick the sequence of actions that reaches a defined goal — think of Google Maps evaluating multiple routes before recommending one.</p><p><strong>Utility-based agents</strong> go further still. Reaching the goal isn’t enough; they want to reach it <em>well</em>. Two routes might both get you to college, but one costs ₹40 and takes 35 minutes while the other costs ₹120 and takes 20 — a utility-based agent weighs cost against time against comfort and picks accordingly, the same mental math you did standing outside your door.</p><p>And finally, <strong>learning agents</strong> — the ones that don’t just plan well once, but get better at planning the more they operate. Every time a recommendation app nails your taste a little more precisely than last month, that’s a learning agent quietly improving its own decision-making.</p><p>The pattern across this spectrum is simple: as agents move from reflex to reasoning, they stop just <em>reacting</em> and start <em>imagining</em> — running a kind of mental simulation before committing to a real-world action.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eOlBi66KcjNp3KW4DvQ7Ew.png" /></figure><h3>Planning Is Just Search, Wearing a Trench Coat</h3><p>Here’s the part that surprised me most while researching this: “planning” in AI isn’t some separate mystical process — it’s search. The agent treats the problem as a graph of possible states, where each state is “a situation I could be in,” and each edge is “an action that gets me from one situation to another.” Planning is just finding a good path through that graph from the current state to the goal state.</p><p>This is the same intuition behind uninformed search strategies like BFS and DFS, which explore states without any sense of direction, and informed strategies like A*, which use a heuristic — an educated guess about how promising a state looks — to explore smart, not just wide. When the world involves an opponent working against you, like in chess, agents switch to adversarial search techniques like Minimax, essentially planning several moves ahead while assuming the opponent is also planning against them.</p><p>To make this concrete, here’s a small agent solving a grid navigation problem with A* search, using Manhattan distance as its heuristic — it’s the same core logic a delivery robot uses to find its way around obstacles on a warehouse floor.</p><pre>import heapq<br><br>def heuristic(a, b):<br>    # Manhattan distance: a quick, admissible estimate of remaining distance<br>    return abs(a[0] - b[0]) + abs(a[1] - b[1])<br><br>def a_star(grid, start, goal):<br>    rows, cols = len(grid), len(grid[0])<br>    open_set = [(0, start)]<br>    came_from = {}<br>    g_score = {start: 0}<br><br>    while open_set:<br>        _, current = heapq.heappop(open_set)<br><br>        if current == goal:<br>            path = [current]<br>            while current in came_from:<br>                current = came_from[current]<br>                path.append(current)<br>            return path[::-1]<br><br>        r, c = current<br>        for n in [(r+1, c), (r-1, c), (r, c+1), (r, c-1)]:<br>            nr, nc = n<br>            if 0 &lt;= nr &lt; rows and 0 &lt;= nc &lt; cols and grid[nr][nc] != 1:<br>                tentative_g = g_score[current] + 1<br>                if n not in g_score or tentative_g &lt; g_score[n]:<br>                    came_from[n] = current<br>                    g_score[n] = tentative_g<br>                    f_score = tentative_g + heuristic(n, goal)<br>                    heapq.heappush(open_set, (f_score, n))<br><br>    return None  # no reachable path<br><br>grid = [<br>    [0, 0, 0, 0],<br>    [0, 1, 1, 0],<br>    [0, 0, 0, 0],<br>    [1, 1, 0, 0]<br>]<br>print(a_star(grid, (0, 0), (3, 3)))</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-3FpzWBSqQc4fXHlTXgpRQ.png" /></figure><p>What makes this “planning” rather than just pathfinding is the <em>why</em> behind every step: the agent isn’t asking “can I move here?” — it’s constantly asking “does moving here get me meaningfully closer to my goal, given everything I currently know?” Scale that idea up from a 4x4 grid to a warehouse with hundreds of robots, or a road network with millions of intersections, and you have the backbone of real-world AI navigation systems.</p><h3>Planning When the Future Refuses to Cooperate</h3><p>Grids and chessboards are tidy. Real life isn’t. Traffic appears out of nowhere, a hallway that was clear five minutes ago is suddenly blocked, a stock price swings on a rumor no algorithm saw coming. Agents operating in the real world need to plan under <strong>uncertainty</strong> — and this is where two ideas take over from simple search.</p><p><strong>Markov Decision Processes (MDPs)</strong> let an agent reason not about one fixed path, but about a tree of possible outcomes, each with a probability attached. Instead of “the fastest route,” the agent computes “the route with the highest <em>expected</em> value, accounting for the chance of delays.” <strong>Utility theory</strong> gives the agent a way to score outcomes that aren’t just right or wrong, but better or worse along multiple dimensions — speed, safety, cost — much like you weighing “faster but riskier” against “slower but safer” without consciously doing the math.</p><p>Push this further and you land on <strong>reinforcement learning</strong>, where an agent doesn’t get handed a perfect model of the world upfront — it learns good decisions purely through trial, error, and reward, refining its strategy every time it acts. It’s less like following a map and more like the way you personally learned exactly how much buffer time to leave for a college bus that’s late roughly one day in three — nobody told you that number, you learned it by getting stuck a few times.</p><h3>You Were the Reference Implementation All Along</h3><p>Every framework above has a human parallel, and honestly, that’s the most useful way to internalize it:</p><p>Your goal state was “reach college on time.” Your possible actions were “walk, auto, bus, train.” Your heuristic was an instinctive sense of which option <em>felt</em> fastest, without calculating exact minutes. Your uncertainty was the rain that might or might not slow the bus down. And your utility function silently balanced time, cost, and comfort — not optimizing for speed alone, because nobody actually wants the fastest possible commute if it means standing in the rain.</p><p>An intelligent agent doesn’t do anything mysterious — it takes exactly this process and turns it into math and code, so that a system with none of your lived intuition can still make a reasonable call in a changing, uncertain world.</p><h3>Where This Actually Shows Up</h3><p>This isn’t confined to textbooks or lab assignments. The same perceive-plan-act loop runs quietly behind a lot of everyday technology:</p><p>Navigation apps replanning your route the moment traffic data updates, in real time, without you asking. Warehouse robots — the kind used in Amazon fulfillment centers — planning collision-free paths among hundreds of other robots simultaneously, replanning on the fly whenever one gets blocked. Game-playing AI, from chess engines to AlphaGo, simulating millions of future positions before committing to a single move. Autonomous vehicles deciding, in a fraction of a second, whether to brake, swerve, or hold course, using exactly the search-and-utility logic described above, just running at a speed no human could match.</p><p>The common thread across all of them: perceive the environment, build or update an internal model, plan several steps ahead, act, and immediately start replanning based on what changed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mHOjYAP7_pZYXnVOZm6y_w.png" /></figure><h3>It’s Not Magic — It’s Just Faster</h3><p>Intelligent agents aren’t intelligent because they know more than we do. They’re intelligent because they do, formally and relentlessly, what we already do by instinct: sense what’s around us, imagine a handful of possible futures, weigh them against what we actually value, and act. Search algorithms give agents a structured way to explore those futures. Utility theory and MDPs give them a way to judge which future is actually worth choosing when nothing is guaranteed.</p><p>So the next time your GPS silently reroutes you around a traffic jam, or a chess engine sees a checkmate ten moves before you do — know that it isn’t doing anything you don’t already do every single morning deciding how to get to college. It’s just doing it faster, and without ever needing a second cup of chai to think it through.</p><h3>References</h3><p>[1] S. J. Russell and P. Norvig, <em>Artificial Intelligence: A Modern Approach</em>, 4th ed. Upper Saddle River, NJ, USA: Pearson, 2020.</p><p>[2] P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,” <em>IEEE Transactions on Systems Science and Cybernetics</em>, vol. 4, no. 2, pp. 100–107, Jul. 1968, doi: 10.1109/TSSC.1968.300136.</p><p>[3] R. S. Sutton and A. G. Barto, <em>Reinforcement Learning: An Introduction</em>, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.</p><p>[4] R. Bellman, “The theory of dynamic programming,” <em>Bulletin of the American Mathematical Society</em>, vol. 60, no. 6, pp. 503–515, Jul. 1954.</p><p>[5] D. Silver et al., “Mastering the game of Go with deep neural networks and tree search,” <em>Nature</em>, vol. 529, no. 7587, pp. 484–489, Jan. 2016, doi: 10.1038/nature16961.</p><p>[6] D. González, J. Pérez, V. Milanés, and F. Nashashibi, “A review of motion planning techniques for automated vehicles,” <em>IEEE Transactions on Intelligent Transportation Systems</em>, vol. 17, no. 4, pp. 1135–1145, Apr. 2016, doi: 10.1109/TITS.2015.2498841.</p><p>[7] J. von Neumann and O. Morgenstern, <em>Theory of Games and Economic Behavior</em>. Princeton, NJ, USA: Princeton Univ. Press, 1944.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b79af667940f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>