Budgets, Stopping, and Human Intervention

Goal

Set hard limits for the whole collaboration system and every subtask. Make completion, failure, pause, cancellation, and human approval explicit states.

A budget is more than money

Consider at least seven limits:

• maximum model and tool calls;
• token or monetary budget;
• wall-clock deadline;
• maximum concurrency;
• maximum delegation depth and subtask count;
• maximum attempts or rework per task;
• quotas for files, network, databases, and other resources.

The global budget covers all sub-agents; do not let each one believe it owns the full allowance. Preserve capacity for recovery and aggregation — a research phase must not consume the writer’s and reviewer’s budget.

Terminal states should be exclusive and explainable

Distinguish:

• succeeded — acceptance evidence is complete;
• failed — completion is impossible, with a stable failure class;
• denied — policy rejected the action;
• budget_exhausted — a hard limit was reached;
• canceled — a user or superior layer canceled actively;
• waiting_human — paused for approval or clarification;
• partial — the contract allows a partial result and names the gap.

“The agent says it finished” must not transition state to succeeded. The scheduler checks all required tasks, output schema, and acceptor results.

Stopping conditions

Beyond success and budget, detect:

• consecutive identical actions or messages;
• rework loops with no new evidence;
• a quality score that does not improve across several rounds;
• a dependency graph that cannot make progress;
• a missing waiting object or expired lease;
• conflict count over a threshold;
• a safety-policy trigger.

After stopping, retain the last stable checkpoint, verified artifacts, reason for incompletion, and recommended human action. Do not automatically widen permission or remove guardrails merely to “continue.”

Human in the loop is not one prompt

An effective approval contains:

• the exact action requested;
• target resource and impact scope;
• a preview of important input and output;
• risk and reversibility;
• who may approve;
• transitions for approve, reject, modify, and timeout;
• the approval record and corresponding task or tool-call ID.

As of 2026-07-22, the OpenAI Agents SDK Human-in-the-Loop documentation describes interrupting a tool call, serializing run state, and resuming after approval or rejection. Interfaces will change, but the engineering principle is stable: pause without executing, let an external principal decide, and resume from the same state.

The SDK aggregates interruptions in an outer RunState, including tools in a handoff or nested Agent.as_tool(), and can serialize and resume it. This is SDK behavior, not a guarantee that external authorization stays valid. Serialized state can include application context and runtime metadata, so govern it as persistent data: do not place secrets or broad capability tokens in context; production approval still requires independent identity, scope, expiration, and audit checks at the execution boundary. See Human-in-the-loop (accessed 2026-07-22).

Actions that normally need a human decision

• External sending, publishing, payment, deletion, or irreversible write;
• data-access or permission expansion;
• high-impact health, legal, or financial decisions;
• conflicting evidence where automatic arbitration is risky;
• user-intent ambiguity that would change the outcome;
• material budget expansion.

Low-risk, reversible, fully validated repetition may be automatically approved by policy, but the authorization must remain revocable.

Budget-allocation example

Total call limit 20
  Manager and routing: 4
  Two parallel specialists: 5 each
  Aggregation: 3
  Review or correction: 3
Total deadline 60 seconds; each tool 10 seconds; maximum concurrency 2
Delegation depth 1; at most 2 attempts per task

If one specialist uses all five calls early, it must not consume the review budget. The manager can deliver partially or stop.

Exercise and self-check

List budgets and at least two approval points for “generate and publish a weekly report.” Does approval timeout mean failure or waiting_human? After user cancellation, how do already-started parallel agents and tools receive the cancellation signal?

Next step

Continue with Evaluation, Observability, and Security Boundaries.

References

• OpenAI Agents SDK: Human-in-the-loop — accessed 2026-07-22.
• OpenAI Agents SDK: Running agents — accessed 2026-07-22.