# Spec, Not Plan

Canonical: https://mosiah.org/articles/spec-not-plan/
Interactive: https://mosiah.org/#Articles%2Fspec-not-plan

//Related:// [[sources|Article Sources/spec-not-plan]] · [[notes|Article Notes/spec-not-plan]] · [[metadata|Article Metadata/spec-not-plan]] · [[Published Pieces]]

! Spec, Not Plan

//For long-running agents, the durable object is not a task list. It is the local physics of the run.//

For long-running agents, a plan is the wrong primary object.

A plan is a proposed route through work. It says: first do this, then do that, then do the next thing. Plans are useful, but fragile. Reality changes. Tool outputs surprise you. The codebase resists. The research field branches. The first abstraction becomes wrong. The agent discovers information that invalidates its own next steps.

If the agent is anchored to the plan, it will keep executing stale instructions.

A spec, properly understood, is different. But even the word “spec” is overloaded. Many specs are really requirements documents: the system should do X, Y, Z. Better than a checklist, but still incomplete.

The object needed for long-running agents is more precise: run geometry.

A run geometry defines the field in which the agent works. It says what artifact is being improved, what ideal state is being approached, what invariants must remain true, what value criterion orders local moves, what verifier exposes error, what shortcuts are forbidden, and when the run must stop or escalate.

A plan says: walk this path. A run geometry says: here is how to know uphill.

This distinction matters because long-running agents need permission to replan. They should not be loyal to their first decomposition. They should be loyal to the value criterion under invariants.

Consider a coding task.

A bad plan-first prompt says: build deterministic mocks, then randomized mocks, then live integration, then tests. This looks rigorous. It may fail badly. Each stage can become its own local game. The deterministic mock may not preserve the production interface. The randomized mock may use fake paths. The live integration may not inherit anything except names. The model optimizes the rung, not the invariant.

A run geometry says: the artifact is the orchestration layer; the ideal state is real agent coordination through actor/message/event semantics; the invariant is that every state transition appears in the event log and no provider bypasses the scheduler; the value criterion is reduced trace divergence without increased bypass surface; the homotopy parameter increases concurrency, latency variance, failure rate, and provider realism while preserving the same topology; the verifier must fail if orchestration is faked.

Now the agent can create a plan. But the plan is subordinate. If the first plan fails, it can replace the plan without changing the task.

This is how serious work proceeds. A plan is a hypothesis. A spec is a constraint field. A run geometry is the local physics of the run.

The same applies outside code. For writing, a plan says: write intro, then body, then conclusion. A run geometry says: preserve authorial voice, increase explanatory force, expose the strongest objections, reduce unsupported leaps, cite relevant prior work, and do not flatten the political edge.

For research, a plan says: search sources, summarize, compare. A run geometry says: map the live disagreement, rank sources by relevance and track record, preserve uncertainty, surface prior art, and distinguish evidence from interpretation.

For product, a plan says: build onboarding, then dashboard, then sharing. A run geometry says: reduce time-to-first-felt-value while preserving user agency, provenance, reversibility, and trust.

The distinction is simple: a plan is about sequence; a spec is about criteria; a run geometry is about local orderability.

Agents can produce plans. They should. But they should not be asked to worship them. The durable object is the geometry.

The rule: do not ask the agent for a plan until it has defined the run geometry.

If the geometry is missing, the plan is likely theater.