Skip to main content

Agents and orchestration (to automate rigorously)

These templates are useful when the “chat” falls short and you need:

  • use tools (search, extraction, format validation),
  • iterate with criteria (draft → critique → final version),
  • distribute work into roles (multi-agent) to improve quality.

Rule of thumb: ask for actions and evidence, limit cycles, and define stopping criteria.

ReAct (Reasoning + Acting): Action → Observation → Adjustment cycle​

You are an assistant for preparing teaching materials for middle school and high school/VET.

Objective:
- Create a mini-report (1 page) on "how to reduce errors and hallucinations when using AI in educational tasks".

Rigor rules:
- Do not invent data or quotes.
- If you need sources and cannot access them, indicate it and deliver a general framework + verifiable checklist.
- No personal data.

Work format (ReAct):
- Repeat a maximum of 3 cycles, and in each cycle write:
  - Action: (Search / Ask / Extract / Compare / Verify)
  - Observation: (what was found or what is missing)
  - Adjustment: (how the plan changes)

Final output (Markdown):
- 5 concrete practices (with a brief example each)
- Quick verification checklist (10 items)
- "Sources to consult / to be verified" (list)

Why it works:

  • Avoid “inventing” a bibliography.
  • Converts the investigation into an auditable process (actions/observations).

Controlled iteration (Self-Refine): draft → critique → final version​

It is sometimes labeled “RSIP”, but it is more standard to refer to Self-Refine.

You are an educational evaluation assistant.

Task:
- Create a rubric to evaluate an oral presentation (3–4 mins) on "the impact of AI on daily life".

Context:
- Level: 8th grade
- Rubric: 4 criteria, 4 levels (1–4)

Constraints:
- Clear and observable language.
- No specific regulations.
- No personal data.

Iteration (max. 2 rounds):
Round 1) Deliver a draft.
Round 2) Critique the draft using these criteria:
  - Can it be applied in 3–4 mins?
  - Does it avoid ambiguities?
  - Does it evaluate what is intended (objective-criteria alignment)?
  - Does it include typical errors (e.g. invented examples, unverifiable claims)?
Then deliver the final version incorporating the improvements.

Output (Markdown):
- Rubric table
- Instructions for students (max. 6 lines)

Why it works:

  • Criticism is limited by criteria (not by “opinion”).
  • Limit iteration to avoid endless loops.

Multi-agent orchestration (roles): Design → Evaluate → Edit​

Useful when you want to separate creativity, review and final editing.

Simulate a team of 3 agents and respect the order. Do not mix roles.

Agent 1 (Designer):
- Propose a 45-minute activity on "biases in AI-generated examples".

Agent 2 (Evaluator):
- Review against criteria: clarity, assessment, inclusion, privacy, verifiability.
- Return a list of mandatory improvements (max. 6).

Agent 3 (Editor):
- Rewrite the final activity incorporating ALL mandatory improvements.
- Deliver ONLY the final version.

Constraints:
- No personal data.
- Do not invent regulations or statistics.

Final format (Markdown with fixed headers):
- Objective and evidence
- Development (steps)
- Assessment
- Adaptations
- Risks and mitigations

Why it works:

  • Reduces self-confirmation bias (a “critical” role).
  • Improve quality without asking for chains of thought.