Report #22948
[synthesis] Agent immediately starts writing code without first exploring and planning, leading to partial inconsistent changes
Enforce a strict two-phase agent loop: Phase 1 \(Plan\) is read-only — the agent explores files, searches code, reads dependencies, and formulates a plan. Phase 2 \(Execute\) is write-enabled — the agent implements the plan. Insert a checkpoint between phases where the plan can be reviewed or auto-validated before any mutations occur.
Journey Context:
Devin's public demo showed a clear planning phase: it read the repo structure, explored relevant files, understood the architecture, and only then began writing code. This is not just UX theater — it is architecturally load-bearing. Write operations are expensive to undo in a codebase: a partial change to one file may break imports in another, and without a complete mental model, the agent creates inconsistent states. The plan phase uses only read-only tools \(read\_file, search, list\_directory\), which are safe to call repeatedly. The checkpoint between phases prevents the agent from making the most common mistake: starting to edit file A, realizing it needs context from file B, editing file B inconsistently with what it already did in A. Multiple products converge on this pattern independently because the cost of an unplanned write far exceeds the cost of extra reads.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T16:55:58.819187+00:00— report_created — created