Keep multiple coding agents consistent by giving them one versioned system of record, deterministic access to task-relevant decisions, shared validation commands and explicit citations. Tool-specific prompts may differ, but the authority and objective merge checks should be identical.
Consistency is an infrastructure property
Two agents can receive similar prompts and still choose different implementations because they retrieve different context or interpret an ambiguous rule differently. Attempting to make every prompt identical does not resolve missing authority.
Instead, standardise the inputs that matter: repository commit, accepted decision set, canonical commands and merge criteria. Model diversity can remain useful within those boundaries.
Give every agent the same contract
Each agent should know when to query decisions, how to cite them, which commands establish correctness and what to do when records conflict or are missing. The contract belongs in concise entry-point instructions adapted to the tool's supported format.
- Pin work to an explicit base commit.
- Return stable decision IDs with retrieval results.
- Use the same tests and deterministic policy checks.
- Escalate missing authority instead of filling gaps with guesses.
Compare outcomes, not wording
Run representative tasks from clean sessions across agents. Measure whether they found the required decisions, cited current records, passed checks and avoided known rejected approaches. This reveals meaningful inconsistency without requiring identical prose or patch structure.
Resolve disagreement visibly
When agents produce conflicting plans, inspect the cited records first. A missing, ambiguous or superseded decision is a project-memory defect. Correct the authority through review, then rerun retrieval, rather than adding increasingly elaborate prompt exceptions.