Skip to content

4. Agent draft loop

About two minutes

This is how agents should use the tool day to day: not one-shot generation, but a small loop that stops when the corpus has nothing more useful to say.

flowchart TD
  A["Phase A: orient on intent"] --> B["Agent drafts damlCode"]
  B --> C["Phase B without compile"]
  C -->|fail| D["Revise from strong patterns"]
  D --> C
  C -->|ready_to_compile| E["Client: daml build"]
  E --> F["Phase B with compilationResult"]
  F -->|pass| G["Done"]
  F -->|fail with diagnostics| D

Revise using retrieved patterns. If two failed revisions have nearly the same alignmentScore (within about 0.05) and no new useful pattern symbols appear, stop asking this corpus for more guidance — the draft may still be wrong. Hand the unresolved issue to a person; compile only when the outcome is ready_to_compile.

When to stop (“as good as the corpus allows”)

Stop and hand off when any of these is true:

  1. Outcome is ready_to_compile - next step is compile
  2. Outcome is pass - done for this loop
  3. Plateau - two consecutive fails with alignmentScore within about 0.05 and no new useful pattern symbols
  4. Cap - for example four Phase B turns. Do not thrash.

What to give the human on handoff

  1. Final damlCode
  2. Last outcome, alignmentScore, and reasoning (fixed server text from score band and compile status — not an LLM line-by-line code review)
  3. Corpus anchors (path / sourceUrl) and any issues
  4. A clear ask: compile, or paste damlc diagnostics back into Phase B

Worked script in the private repo: docs/internal/DEMO_CHAT.md Part 2.

← Phase B · Try it →