Falk

Learning & Feedback

The agent does not learn from conversations. It improves when humans update config files based on feedback and evals.

Feedback loop

User asks question
     ↓
Agent answers → trace in Logfire
     ↓
User reacts 👍 or 👎 → score in Logfire
     ↓
Data steward reviews low scores
     ↓
Updates config/context files → agent improves

The improvement cycle

  1. Find issues — filter Logfire traces by low scores
  2. Understand why — see the full trace (query → tools → response)
  3. Fix the source — update synonyms, gotchas, rules, or context
  4. Write a test — add a case to evals/ to prevent regression
  5. Verifyfalk test

Everything is files

All agent knowledge lives in version-controlled files. No database. No migrations. PR-reviewed and version-controlled.

See also