Require evidence before granting agent autonomy

Learning objective

Justify autonomous agents, or bounded agent steps inside workflows, only when unpredictable steps demand them and the quality gain outweighs latency, cost, and compounding-error risk.

Core idea

Recognizing an unpredictable path is the starting point, rather than sufficient justification for autonomy. You must also show that letting Claude choose the path produces a worthwhile result within the team’s latency, spend, and risk constraints.

Anthropic recommends adding complexity only when it demonstrably improves outcomes. Its agent guidance identifies both the benefit of flexible, model-directed execution and the costs of extended execution and compounding errors. The professional decision is therefore conditional: grant discretion to the smallest scope where a simpler path fails and measured gains justify the operating exposure.

Write the justification against a baseline. “The agent completed the demo” does not establish that a focused augmented call, a workflow, or staff triage would perform worse. A defensible autonomy gate compares those alternatives on the same task and preserves the failures that the extra discretion introduces.

How it works

Separate unknown evidence from unknown steps

A task can have missing input while retaining a predictable path. If every incomplete application needs the same missing-field request, the uncertainty is in the values. Code can keep the process fixed. An agent is more useful when one finding changes which evidence or action is necessary next.

Document a representative path the team cannot prescribe. For an incident investigation, a failed request might reveal a configuration mismatch, which points to an unexpected dependency, which determines the next test. Contrast that with a mandatory checklist whose items stay the same across incidents.

Then test whether prepared evidence removes the uncertainty. Anthropic’s context-engineering guidance notes that runtime exploration is slower than precomputed retrieval and supports hybrids that supply some data first. Give the investigator the known starting evidence before paying for it to rediscover that material. Preserve exploration for the information whose relevance emerges during execution.

Measure the gain that matters

Compare a bounded agent with the best plausible simpler design, using the same representative cases and acceptance criteria. Anthropic’s building-block guidance calls for performance measurement and iteration; here the acceptance artifact is a comparison of accepted results, unresolved cases, harmful conclusions, elapsed time, and estimated spend.

Measure quality at the business outcome. An investigation that produces more findings but no reproducible explanation has not necessarily improved delivery. Include cases where an early mistaken premise leads to later irrelevant searches or an incorrect final diagnosis. Those cases expose the compounding-error risk that a successful run hides.

Separate request classes. An agent may earn its cost on exceptions while providing no gain on routine cases. If only the exception class benefits, use a bounded step on that branch rather than expanding autonomy across the whole intake process.

State the team’s acceptance thresholds before comparison. They are workload requirements, not Anthropic performance guarantees. Avoid defending autonomy with vendor benchmark improvements from another task or with an average that conceals cases missing the deadline.

Bound the granted scope

Anthropic recommends sandboxed testing with guardrails and environmental feedback. The production counterpart is a permitted tool set, a concrete completion check, and enforced exits. A tool’s observed result can refute an assumption; a generated plan alone cannot.

Use the Agent SDK’s TypeScript Options.maxTurns and maxBudgetUsd as execution safeguards, and send limit-exhausted work to the declared incomplete-work path.

Treat the spend threshold as an estimated stopping mechanism. The SDK compares spend after model responses arrive, so the response that reaches the threshold still completes and can take spend past it; subagents still running at that point can also spend until they stop. Leave headroom instead of presenting maxBudgetUsd as an exact billing ceiling. Neither a turn limit nor a spend threshold establishes a wall-clock deadline. Pass an abortController in TypeScript Options so the host can cancel operations when its deadline arrives.

Preserve the observations and the stopping reason for the receiving owner, rather than accepting the last plausible draft.

Keep mandatory acceptance outside exploration

A bounded agent step has an entry contract and an exit contract. The entry states the goal and available evidence. The exit carries the conclusion, supporting observations, and unresolved blockers to the outer workflow. The surrounding process then validates and approves the result. Also define a mid-run human checkpoint that pauses the investigator when a finding needs judgment, before it continues exploring; Anthropic describes checkpoints separately from blockers and completion.

A bounded agent step can pause at a checkpoint and continue exploring, and it ends either through its exit contract to the outer workflow or on the incomplete-work path.

Enforce per-call policy with a blocking PreToolUse hook (see 1.2.2). Put limits and policy enforcement where execution occurs, rather than asking the investigator to remember them.

The distinction matters when a workflow has an irreversible final action. Letting an investigator search freely within a narrow scope does not require giving it authority to publish, deploy, or update a business record. A fixed acceptance stage can retain that authority even when the preceding path varies.

Key distinctions and trade-offs

Path flexibility versus output variability: Different answers do not imply different necessary actions. Require evidence that discoveries change the path before adopting autonomous execution.

Full-process autonomy versus a bounded step: Delegate the investigation when its actions are unpredictable; retain prescribed stages whose order and acceptance requirements are known. A hybrid limits how far investigative errors can travel.

Limit exhaustion versus completion: The limit prevents further execution. It does not certify the result. The fallback owner needs the observations already gathered and the reason the run stopped.

More exploration can improve coverage while also creating more opportunities to follow a false lead. Grounding every decision in tool feedback and checking the final artifact help contain that risk, but do not eliminate the need for comparison against a simpler baseline.

Architecture / decision scenario

A technology company’s on-call team has twenty minutes to produce a reviewed incident diagnosis. Automated work must remain within a team-set spend allowance, and production changes require an engineer’s approval.

A fixed diagnostic workflow retrieves the usual logs and checks known failure signatures. It handles recurring incidents well but leaves unfamiliar dependency failures unresolved. The team tests a bounded investigator on both classes, scoring reproducible diagnoses and misleading conclusions while recording elapsed time and estimated spend.

Routine cases retain the workflow because autonomous exploration adds no useful acceptance gain. On unfamiliar cases, the agent can follow newly discovered dependencies and produce evidence that the fixed checklist misses. That measured improvement justifies an investigation branch, subject to the twenty-minute operating constraint.

The TypeScript host sets maxTurns and maxBudgetUsd with spend headroom, passes an abortController that it aborts at the twenty-minute deadline, scopes built-in investigation capabilities with an explicit tools array, and uses a blocking PreToolUse hook for prohibited operations. If execution stops before acceptance, the host returns the evidence to the on-call engineer as incomplete. A finding that needs the engineer’s judgment pauses at a mid-run checkpoint before further investigation. The outer workflow checks the diagnosis and requires approval before remediation.

The team rejects autonomy over the whole incident process. The necessary gain comes from unpredictable evidence gathering; changing the production-approval sequence contributes no measured benefit and increases exposure. The justification records that limited grant and the evidence that could warrant revisiting it.

Common exam traps

• Granting autonomy because inputs are incomplete or the task is long.
• Comparing an agent with an intentionally weak baseline.
• Using more findings as a substitute for accepted business outcomes.
• Giving every request the exception path’s autonomy.
• Treating an estimated spend threshold as an exact billing cap.
• Accepting limit exhaustion as successful completion.
• Letting the investigator decide whether required approval happens.

What to remember

Unknown steps establish a candidate use for autonomy. Measured quality gains and bounded exposure establish the justification. Supply known evidence, delegate the uncertain investigation, enforce limits and permissions, and return incomplete work to a named owner.

Knowledge check

References

• Anthropic, Building effective agents
• Anthropic, Effective context engineering for AI agents
• Anthropic, Agent loop
• Anthropic, Control execution with hooks
• Anthropic, TypeScript SDK reference