Continuing is not the same as progressing

An agent can always produce another plan, call another search tool or try a variation. None of those acts proves that the next step is useful. A system demonstrates judgment when it can connect another action to missing evidence or a measurable improvement—and stop when that connection disappears.

The OpenSkool agent loop ends with CONTINUE / STOP because stopping is a decision, not an absence of capability. Success, uncertainty, policy, cost and human authority can all provide legitimate endings.

Appropriate autonomy changes by action

A research agent may search several public sources without interruption, pause when evidence conflicts and ask before sharing a conclusion externally. A scheduling agent may read availability, draft an event and wait before inviting people. The same system can have wide freedom to analyse and narrow authority to act.

CAPABILITY, PERMISSION, COST and HUMAN OVERSIGHT form the boundary around the loop. Removing the person does not remove the need for that boundary; it moves responsibility into less visible assumptions and code.

Evidence & context: Model Context Protocol · NIST

Too many stops can make an agent useless

There is a real counterargument. A system that asks for approval at every harmless step transfers the work back to the user and creates approval fatigue. Excessive caution can hide weak tool design and prevent the adaptability that justified an agent.

The answer is not a universal confirmation prompt. It is careful allocation: let safe, reversible and well-tested actions proceed within limits; require people where authority, consequence or genuine uncertainty begins.

Evidence & context: Anthropic

Judge an agent by its boundaries

  • It acts when the goal, permission and next step are clear.
  • It verifies environmental change rather than narrating success.
  • It asks when essential information is missing.
  • It escalates when judgment or authority belongs to a person.
  • It stops on success, policy, budget, repeated failure or exhausted evidence.

A demonstration celebrates how far an agent can go. A responsible production system also shows where it will not go. Appropriate autonomy makes that limit part of the design and part of the evidence by which the system is judged.

Evidence & context: Anthropic

Sources & further reading

  1. Building effective agents

    Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.

  2. Model Context Protocol tools

    Model Context Protocol. The current official tools specification checked 13 September 2026. Draft details can change; OpenSkool relies on the durable separation between tool discovery, model selection and host-controlled execution.

  3. Demystifying evals for AI agents

    Anthropic. A provider's engineering guidance on multi-turn agent evaluation, checked 13 September 2026. Examples inform evaluation design but do not establish universal pass thresholds.

  4. Generative Artificial Intelligence Profile (NIST AI 600-1)

    NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗