Self Correction

Agents produce errors: wrong calculations, hallucinated facts, malformed outputs, incomplete task execution. Self-correction catches these before the user sees them — the difference between a reliable production agent and a demo that works 60% of the time.

Self-correction is spell-check for agent outputs — catching errors automatically before anyone else sees them, and suggesting fixes.

Scenario

Agent calculates '15% of $2,400 = $360.' Validator: 2400 × 0.15 = 360 ✓. Critic: 'Answer is correct but user asked for monthly, not total — $360 is annual.' Corrector revises: '$30/month ($360/year).'

Solution

In Agent Foundations, apply Self Correction to this scenario: Agent calculates '15% of $2,400 = $360. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Self Correction (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Self Correction happens in the code.

1def self_correcting_agent(task: str, max_corrections: int = 2) -> str:  # define a reusable function2    result = agent_run(task)3
4    for attempt in range(max_corrections + 1):  # key line for Self Correction5        validation = programmatic_check(result, task)6        if not validation["pass"]:7            result = corrector_llm(task, result, validation["issues"])8            continue9
10        critique = critic_llm(task, result)11        if critique["accept"]:12            return result  # return the result13
14        result = corrector_llm(task, result, critique["issues"])15
16    return result  # best effort after max corrections

• No validation — errors reach the user
• Only LLM critic — slow and expensive for simple checks
• Infinite correction loops — need max attempts
• Not cross-checking claims against tool results
• High correction rate ignored — signals systemic prompt issues

Quick recap — the most important points from this module.