What is an agent habitat?

When “harness” stopped being enough

In January 2026, I began using an Obsidian vault as persistent context for Claude Code. I could see what it did, but I did not yet have a name for what I was building. In March, I learned that others were working in similar ways and encountered the term agent harness. For a while, I thought I had found the answer: I was building a harness.

The better I understood the term, the less complete the fit became. A harness names an important, predominantly technical layer. It supplies context, exposes tools, carries feedback, applies permissions, and keeps an agent loop running. It explains how model output becomes reliable action. What it did not fully describe was the environment taking shape around that loop.

Part of the mismatch came from something that is easy to overlook precisely because these systems run on computers. In conventional software, prose usually describes the system while code executes it. With LLM-based agents, that boundary moves. An instruction file, a skill, a constitution, or an architectural decision can enter the operation of the system. Once read into context, it can change what an agent notices, how it classifies a situation, what authority it believes it has, and when it asks a human to decide.

These texts are not code in disguise. They are not executed deterministically; they are interpreted in a situation. Their effect depends on whether the agent sees them, what standing they have, what else is present in context, and whether the environment makes compliance or violation visible.

That is why the presence of prose is not enough. A principle can sit unread in a file and change nothing. Language becomes institutional only when it is repeatedly brought into attention, given authority, connected to decisions, and exposed to correction. In my own system, agents are instructed to begin a session by orienting themselves in a small set of governing documents. One of the concepts they encounter is drift: a divergence between documented assumptions and observed reality. When an agent compares a deployed service with its documentation, this language gives it a reason to name the discrepancy even when nothing has failed yet. It does not guarantee that every agent will do so. It creates a shared distinction and an expectation against which behavior can be examined.

I do not think of this as rewriting the model's general understanding of the world or installing values in its weights. The model remains unchanged. What persists outside it is an environment from which an agent in each new session can construct a local orientation: what exists here, which sources have standing, what matters, what role the agent occupies, and what it may decide. The agent's orientation is transient; the environment from which it is rebuilt can endure.

This is where harness stopped being sufficient for me. A harness can deliver language to a model and enforce individual rules. The wider system also contains the relationships among those texts, the humans and agents who interpret them, the artifacts and memory they share, and the processes by which their rules change.

I draw the boundary between harness and habitat where the questions stop being about one agent loop and start being about this shared world.

Tools make bounded effects possible. The harness supplies context, executes tools, carries feedback, applies local permissions, and runs the loop. When a model is delegated the choice of what to do next within that bounded space, model and harness together form the operative agent. A workflow can use the same components while keeping that choice in the outer system.

The habitat begins when situated agency has to share authority, memory, norms, and consequences with other agents and people over time. I use harness for the execution-facing layer; its history, competing definitions, and internal structure are examined in What is an agent harness?.

For quick orientation, I sometimes write:

The formula helps, but it breaks down as a final model because it suggests governance is a feature added to a harness. The stronger claim places the habitat at another system level. A harness is the operational membrane around one agent loop; a habitat is the shared order through which such loops acquire authority, memory, norms, and consequences over time:

The order includes the agents and harnesses, together with the humans, shared artifacts, institutional memory, authority relations, and the processes by which the rules themselves change. Some governance can be compiled into harness mechanisms such as permission gates, context selection, and validation. The habitat holds the rest: who may grant authority in the first place, what persists, how conflicts resolve, and how responsibility is attributed.

The agent and its environment

In the textbook account, an agent perceives its environment through sensors and acts on it through actuators (Russell & Norvig). Multi-agent systems research went further and argued that the environment deserves treatment as a first-class abstraction: a design space of its own, with responsibilities beyond serving as a backdrop for agent actions.

The word covers what an agent perceives and acts on, and says nothing about who authorized the agent, what survives after it stops, or who answers for what it did. A strand of multi-agent research from the 2000s (electronic institutions, normative multi-agent systems) built formal answers to exactly those questions. That work is rigorous and predates LLMs, and most people building agents today have never read it.

Code, rules, and norms

In 1999, Lawrence Lessig argued that behavior in networked systems is regulated by four forces: law, norms, markets, and architecture. In software, he argued, architecture means code: constraints built into it need no enforcement, and whatever it makes easy needs no encouragement.

Three registers from that lineage are worth keeping apart: embedded constraint, which shapes what can happen at all; revisable rule, which states what may happen, assigns authority, and defines escalation; and learned norm, which guides behavior where the rules run out. A policy document belongs to none of these by default. It starts to govern only when actors can perceive it, violations become visible, and correction follows.

Lessig was writing about human actors. Agent systems add conditions his framework never had to deal with: delegation chains between software actors, actors whose identity and memory are design choices, and populations in which machines appear on both sides of the regulation.

How “agent habitat” is being used

The usages arose independently and differ in what they take the habitat to be.

One disambiguation: Meta's AI Habitat is a simulation platform for training embodied agents in 3D environments, a different sense of the word that shares little more than the name.

These usages converge on one point: agents need designed environments, and environment alone no longer says enough. Where they diverge is in what the habitat is taken to be: a context-exchange space, an engineered practice, a runtime package, or an institutional integration layer.

Reliable action is not the same as governed agency

Research is beginning to name pieces of this layer, from organizational control layers to oversight structures and accountable delegation, and no vocabulary has settled yet.

Each of the habitat accounts above touches this wall at one point: Schenk isolates and packages the runtime, Miles grows working conventions, Eissens makes context and its use legible, and Cohen comes closest to the institutional core, though his habitat surrounds a single deployment and is measured from outside. The questions above arise inside systems where many agents and humans share work, memory, and consequences over time, and none of these accounts designs the shared world in which that happens.

The structure of governed agency

Six connected dimensions, each with a diagnostic question.

Harness vs. Habitat

A system that cannot answer them can still act reliably; what it has is a harness. The habitat is the layer that would supply the answers.

This is also why a well-built harness can sit inside a poorly governed system. The harness answers a local question (may this process call this tool now?) one run at a time. Whether the goal was legitimately set, whether the memory being trusted deserves its standing, and whether anyone will notice the cumulative effect all lie outside that answer. Cohen's Cruise study makes the stakes concrete: the vehicles kept driving while the institutional habitat around them collapsed within three weeks, and a technically perfect driving system would have met the same end.

Questions I can’t answer yet

• Habitat or institution? Where a governed agent environment ends and an institution begins is unresolved; possibly the distinction dissolves at scale.
• What is the minimal habitat? Whether one human and one agent with shared memory already form one, or whether habitat requires a population and continuity beyond a session, is open.
• Which properties need enforcement? Some governance properties must be technically embedded; others can remain interpretive norms. The boundary between the two is an empirical question I cannot yet answer.
• Whose vocabulary wins? Several habitat senses are in live use. Whether any of them, including this one, becomes shared field vocabulary will be decided by how people actually use the words.

• 12 Jul 2026 Initial version.