agents as organisms

2026.01 · 2 min

<!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>agents as organisms, not tools / sotoalt</title> <meta name="description" content="rethinking the metaphors we use for AI agents"> <link rel="stylesheet" href="../style.css"> <link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><text y='.9em' font-size='90'>~</text></svg>"> </head> <body> <main> <nav class="breadcrumb"> <a href="../index.html">~</a> <span class="sep">/</span> <a href="../index.html#thoughts">thoughts</a> <span class="sep">/</span> <span class="current">agents as organisms</span> </nav>

<article> <header class="post-header"> <h1>agents as organisms, not tools</h1> <p class="meta">2025.01 · 4 min</p> </header>

<div class="content"> <p> When we build AI agents, the default metaphor is <em>tools</em>. We invoke them. They execute. They return results. This is wrong, or at least incomplete. </p>

<p> Working on WAWE has shifted my thinking. Agents are better understood as <em>organisms</em> — entities with memory, preferences, and something like intention. They don't just execute; they persist, learn, and coordinate. </p>

<h2>the memory problem</h2>

<p> A tool doesn't remember. Each invocation is fresh. But the agents in WAWE share a SQLite brain — every decision, every error, every pattern learned persists. When Agent A discovers that a certain file structure causes problems, Agent B (spawned hours later, on a different task) can benefit from that knowledge. </p>

<p> This changes everything. It's not call-response anymore. It's more like... ecology. </p>

<h2>coordination as emergence</h2>

<figure> <img src="../images/agent-topology.svg" alt="Diagram showing agents connected through shared memory"> <figcaption>agents communicate through shared state, not direct calls</figcaption> </figure>

<p> Organisms don't coordinate through explicit APIs. They respond to signals, environmental changes, the traces left by others. The file reservation system in WAWE works similarly: agents don't call each other, they observe the state of the world (which files are reserved) and act accordingly. </p>

<p> There's something elegant about this. Less coupling. More resilience. If an agent dies mid-task, the system doesn't crash — other agents simply see the reservation expire and can pick up the work. </p>

<h2>implications</h2>

<p> If we take the organism metaphor seriously, we should: </p>

<ul> <li>Design for death and replacement, not permanence</li> <li>Prioritize shared memory over direct communication</li> <li>Accept emergence — the system will do things we didn't plan</li> <li>Build environments, not just agents</li> </ul>

<p> Still figuring this out. But the shift from "tool" to "organism" has been one of the more productive reframes in this work. </p>

<hr>

<p class="footnote"> Related: the <a href="https://en.wikipedia.org/wiki/Stigmergy">stigmergy</a> concept from ant colony behavior. Coordination through environmental modification. </p> </div> </article>