The bounded session

A bounded work session is a prepared responsibility agreement. It gives the agent one observable outcome, names what belongs inside and outside the work, plants the decisions that stay human, and states which checks could establish the result. Inside that agreement, ordinary choices may be delegated when current evidence settles them.

That boundary does not make judgment automatic. A plan can be stale. A product mechanism can recognize selected consequential actions, but it cannot recognize every decision hidden inside a change. The agent can also misunderstand evidence or fail to notice that it has reached a seam.

Prepare one finishable problem, delegate the evidence-settled work inside it, and keep product intent and consequential judgment with the human.

[Visual 2.2 — The work-loop chain with the Session block enlarged and the other three dropped to thin outline. Inside the enlarged block, four labelled bands stacked top to bottom: "Outcome — one observable result", "Not in scope — named before starting", "Human decision — planted, not discovered", "Verification — checks that could fail". A dotted line crosses the block's right edge outward to a small parked box labelled "a later session", carrying the tag "discoveries get an owner, not a detour". The Plan block on the left keeps a thin arrow into the top band, labelled "prepared here".]

An ordinary example

Return to the small order app. The next job is to let an account manager export the orders already visible under the page's date and status filters. A vague request such as "improve exports" gives the agent too many jobs at once. It could redesign filtering, add PDF output, introduce a worker queue, migrate old order statuses, or expose fields that happen to exist in the database.

A bounded session says something more precise:

• Outcome. An authorized manager receives a CSV of the currently filtered orders.
• In scope. The existing query, the export endpoint, the CSV serializer, the export control, and the authorization tests.
• Not in scope. Filter redesign, PDF output, background processing, analytics, email delivery.
• Human decision. Which customer fields may leave the application.
• Verification. Authorization, filtering, approved columns, escaping, an empty result, and the existing list behaviour.

The agreement delegates ordinary execution without granting authority over every decision the work might reveal. The agent can inspect the application's current patterns, follow an established serializer approach, implement the endpoint and run the named checks. It does not need a new human decision before every reversible edit when the prepared evidence already settles the choice.

The approved customer fields are different. The database may contain more fields, but their mere existence is not permission to export them. The session plants that decision before execution, so the agent knows the boundary and the human remains responsible for what leaves the application.

Prepare a result, not an ambition

A finishable problem has one coherent outcome that can be observed and checked. It does not have to be one prompt or one feature. An audit, an investigation, a migration phase, a refactor, a research pass or a cleanup can all be finishable. What matters is that the session can say what should become true and where its responsibility ends.

The one-sentence goal is a useful sizing test. If the sentence says "add CSV export and redesign filtering", the two halves probably want different evidence and different finish tests. That is a signal to split them, not a grammatical law. Some coherent jobs contain the word "and", and some oversized jobs hide behind a short sentence.

Preparation should name the outcome, the files or behaviours in scope, the explicit non-goals, the current sources, the important risks, the points where a human decides and the verification. A detailed plan built on stale assumptions is still false preparation. The evidence has to be current, and the checks have to be capable of disagreeing with the plan.

Put human attention at the judgment seams

Continuous supervision spends human attention on reversible mechanics. Fully unguided work has the opposite failure: the agent may invent product intent or make a consequential choice because no boundary told it to stop. A prepared session separates three kinds of seam.

A planted human decision is known before execution. The export-field list is one. It belongs in the session agreement beside the work, not in a later conversation after the agent has already chosen.

A recognizable consequential seam appears in a form the product has been designed to notice. At selected seams, Coldstart can bring the pending action back for a human decision. It can also refuse selected actions that violate the declared kind of session. A clean call produces no such decision; that silence means only that no represented rule matched. It is not proof that the action is harmless or intended.

A contextual judgment seam depends on meaning the mechanism cannot derive. A change to what customers can export may be hidden inside ordinary-looking application code. The plan and the agent's reasoning are responsible for noticing it, and either can fail. That is why the planted decision and the nearby limit matter more than a broad promise that the boundary will always fire.

If the existing query behaves differently from the documented filter contract, the agent first asks whether current evidence settles the difference. Adapting an implementation to a settled contract can remain ordinary work. Changing the exported result is a product and exposure decision, so the human owns it even when no mechanism recognizes the seam.

Give discoveries a durable disposition

Real work rarely follows a plan exactly. An in-scope fact required for the outcome can be handled now. Adjacent work receives a durable future owner. A separable responsibility becomes another session. A missing decision or a failed critical check becomes a blocker. The agent may continue independent work, but it must not build on a guessed answer.

Suppose the agent discovers that a large account cannot finish the export within the agreed synchronous limit. A background export may be worth preserving, but adding a worker, retries, monitoring and new failure modes would change the responsibility and evidence the session was prepared against. The current session can enforce the agreed limit and queue a separate design. It does not quietly become a queue-architecture project.

Verification earns the close

The plan names verification before execution so "the code is written" cannot quietly become the finish condition. For the export, the checks cover authorization, current filters, approved columns, escaping, an empty result and unchanged list behaviour. Each observation has to be specific enough to fail, and an unrun check remains unrun.

Running those checks and deciding whether the result can close is a separate responsibility. The bounded session prepares that close; it does not turn an automatic trace or a conversation summary into proof.

Hand the next session a path, not a replay

After a verified close, a fresh session needs an address back to current state, not a replay of the entire conversation. A short pointer can name the active work, its next action, blockers and a small reading list. The repository and current checks still supply the detailed reality.

That pointer is a later feature with its own job. Here the important limit is simple: a concise handoff can be stale, incomplete or contradicted by concurrent work. It narrows the route back to the evidence; it is not perfect memory.

What this does not claim

Preparation narrows responsibility. It does not guarantee success, eliminate mistakes or make the boundary impossible to miss. The agent can misread evidence, a plan can be stale, a check can be weak, and some judgment cannot be recognized mechanically. The human stays responsible for intent, exposure, consequential scope and irreversible decisions.

Coldstart has not demonstrated a general productivity, quality, or error-rate gain from this method. For a small and obvious change, the planning and closeout ceremony can cost more than it saves. The heavier workflow earns its place only when recurring drift, risky seams, adjacent work, or multi-session handoffs make the boundary valuable.

Next

Running the named checks and deciding whether the work can finish is the next stage: verified close.

How current this page is

This page was checked on 2026-08-10 against Coldstart's current planning, close and risk-boundary behaviour. It describes a responsibility model, not evidence that the method improves outcomes. To ask whether a claim still holds, or to report one that does not, write to [email protected].