Minimum sufficient loop

A useful agent does not seek the longest possible trajectory. It seeks a sufficient one, then stops when further iteration has little expected value. Teams that confuse iteration with diligence pay model cost as if it were care. Teams that stop too early dump cleanup onto humans downstream. Both failures trace to the same gap: nobody designed sufficiency as architecture.

Helpful context: This page assumes familiarity with stop conditions as design, what is an agent, and evaluation. It applies those ideas to loop depth.

Five sufficiency checks

From the Meal Companion grounding reference, the agent stops when all of the following hold. The outcome is useful: a practical weekly plan for the household, not a fluent essay about food. Exact checks pass: allergies, portions, schema validation on endorsed knowledge. Material assumptions are visible: busy evenings, spinach deadline, leftover willingness. Authority is respected: no unsanctioned purchase, no override of confirmed restrictions. Further iteration has little expected value: one local weakness remains versus full replan territory.

Missing any check means pursuit continues, escalates, or asks. Meeting all five means stopping is responsible, not lazy. A familiar, reversible week may need one interpretation, one composition Skill, one critique, and stop. A novel medical restriction returns the work to full deliberation and human leadership. Depth must be earned by the terrain, not by a default iteration budget. The principle also prevents over-iteration on routine cases: when all five checks pass after one lap, continuing to iterate produces diminishing marginal improvement while consuming budget that could serve the next household request.

Apply to week planning

Walk the originating sentence: "Plan dinners for four people this week. Tuesday and Thursday are busy. Use the spinach before Wednesday and avoid meals we ate last week." After composition and critique, one weekday weakness may remain. If local repair via RecommendMealSubstitution resolves it while preserving accepted meals, another full replan likely has low expected value. Stop.

If critique reveals the week representation itself is wrong, local patch wastes budget. Replan. If preference is missing and would change the plan materially, ASK with a targeted question beats another generation pass. The applicability guard on expertise like "Busy Week Pattern" performs the same sufficiency thinking at strategy level. When the guard passes, a compressed path is licensed. When it fails, return to full deliberation.

Ask before another step

Before authorizing another lap, the agent through policy encoded in ordinary code answers several questions. What gap will the next retrieval fill? Which weakness will the next repair address? What uncertainty will the human question resolve? Why is another critique likely to improve the result? When answers are unclear, stopping may be wiser than continuing. Repetition without state change wastes budget without adding evidence.

Failure mode	Pattern
Repetition without state change	Retry with altered wording, same missing evidence
Critique without repair locality	Regenerate everything after one local weakness
Endless improvement	Criteria already met, style keeps changing
Premature acceptance	Plausible artifact without hard constraints or authority checks

Depth must be earned by the terrain.

Introduction to Thoughtware Ch. 20

Caps versus sufficiency

A fixed iteration cap protects cost. It cannot define sufficiency by itself. Caps are guardrails. Sufficiency is judgment about value, risk, and evidence. Product teams sometimes ship caps because stops are hard to specify. That trades one failure for another: silent under-deliverance when the cap hits mid-repair, or over-spend when the cap is high and the loop lacks review discipline. The right sequence is to design stops tied to the five checks first, then add caps second as economic backstop. Log which check triggered stop so evaluation can improve policy over time. In practice, teams that rely solely on caps find that most runs terminate by hitting the budget wall rather than by satisfying quality conditions, which means the system is routinely delivering incomplete work that passes only because the cap provides plausible deniability for premature acceptance.

Skills compress familiar terrain

When the same sufficient path keeps appearing, name it as a Skill and invoke the settled procedure. The minimum sufficient loop then applies to remaining uncertainty, not work the system already knows how to perform. The Busy Week Pattern exists because full deliberation every Tuesday-heavy week had low marginal value once guards and checks were stable. Compression is expertise, not shortcutting safety. Failed guards return the case to full loops.

Pipelines are not agents because they end at procedure completion, not at sufficiency against a goal. Skills may live inside loops as perform steps. Agents own whether another lap is worth its cost.

Connection to trust and consequence

Downstream consequence makes premature acceptance visible. A plan that reads well but violates allergy enforcement fails exact checks even when prose is confident. Recovery is part of trust when a stop triggers escalation rather than silent give-up. Minimum sufficiency is therefore a trust design problem. Households should see material assumptions. Operators should see why the loop stopped. Evaluators should score trajectories, not final text alone.

Terrain changes the loop length

The same agent runs different depth on different weeks. A familiar reversible week with endorsed knowledge and no novel restrictions may satisfy sufficiency quickly. A week with unfamiliar medical language, conflicting constraints, or missing preference should lengthen deliberation or return leadership. Judgment terrain is the honest reason one case needs one lap and another needs five. See applicability conditions and finding the weak decision. Sufficiency without terrain awareness produces either over-looping on easy cases or under-looping on hard ones.

The familiar week walkthrough makes this concrete. Two busy evenings, spinach to use, no novel restrictions, endorsed allergy knowledge loaded. The agent may interpret once, invoke Compose Weekly Plan, run critique, patch one local weakness, and stop. Each sufficiency check has a clear answer. The loop stays short because terrain permits compression. Compare with a week that introduces unfamiliar medical language. Guards fail. Leadership returns to the household. Sufficiency requires more laps or escalation. The same agent runs different depth honestly.

Document stop reasons

Production systems should log which sufficiency check triggered stop, which check failed when pursuit continued, and which transition followed. Those logs become training material for policy refinement and for evaluation suites at trajectory level. Without logged stop reasons, teams debate whether the agent stopped too early based on anecdote. With them, evaluation can ask precise questions: did the loop stop before allergy checks passed, before authority was obtained, before material assumptions were shown?

Behaviour standards influence when asks and challenges fire. Sufficiency checks assume the agent exposed material assumptions and refused silent guessing. A loop that stops while hiding assumptions fails the visibility check even when the plan reads well. Minimum sufficiency gives product and finance shared language. Another lap needs a predicted gap filled, not a vague hope for better wording. Stop when checks pass and marginal value drops. That framing prevents both runaway spend and premature ship.

Sufficiency is measured against the declared goal, not against the most elaborate loop the team could imagine. Extra iterations without new information usually add cost without new judgment.

What to do next

Write the five sufficiency checks for the agent goal in plain language and map each check to code, cognitive unit, or human handoff. Add review questions before another lap is authorized. Log stop reasons and review the agent loop transitions that implement them. Sufficiency is a product decision expressed in architecture, not a model hyperparameter. Tie sufficiency checks to acceptance criteria in the spec, because acceptance criteria without stop checks invite premature ship.

See stop conditions as design, where the cognitive unit ends and the agent begins, and cognitive orchestration.

Read next: The agent loop.