Parallel specialists should read a named immutable snapshot or record the exact version each one observed. When outputs are combined, the controller compares those versions before treating results as one consistent picture. If agents can write, assign non-overlapping ownership of files, records, or actions; use a merge gate with conflict detection. A shared directory does not grant coordination by itself. Two agents editing the same file from the same old base can silently overwrite each other even if both prompts request careful work. Prefer read-only specialist analysis and a single controlled writer for consequential changes. Recheck external state immediately before any effect.
Make this comfortable
Parallel agents: isolate state and assign one writer
Operational case
Telemetry reads incident snapshot PK-214 at 09:26 UTC. Release review accidentally reads PK-211 from 09:09, before release RC-47. Its conclusion that pool size was 32 conflicts with a later configuration record showing 8, but this is a version mismatch, not proof that either agent is dishonest. The controller re-runs the release packet on PK-214 before merging. In a separate proposed remediation, two agents both want to edit the same deployment manifest. The controller assigns one writer after analysis and requires a fresh diff against current main. No parallel agent is allowed to commit a competing manifest patch.
Common read snapshot: PK-214.
Telemetry result: PK-214; metrics 09:17..09:26 UTC.
Release result from PK-211: reject for merge; rerun on PK-214.
Write ownership: deployment manifest has one assigned editor.
Before effect: compare current version with reviewed patch base.Performance and operating cost
Version checks across K results are O(K); a file-level merge can be much more expensive if overlapping edits require human resolution. Parallel read-only work often reduces wall time, but snapshot creation, storage, and reconciliation add cost. A single writer may serialize the final edit; that is acceptable when it prevents lost updates. The prompt should name ownership, while optimistic version checks or locks in the application enforce it. Keep a conflict as an explicit unresolved state rather than asking a model to blend incompatible records into a smooth narrative.
Common Mistakes
- Do not combine findings from different state versions without reconciliation.
- Do not let two agents write the same resource because their descriptions sound independent.
- Do not treat a shared workspace as a locking or version-control protocol.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Conversation checkpoints: resume from verified state
- Tool plans: separate independent reads from dependent actions
- Workbook edits: review a cell diff before applying changes
- Specialist agents: prove that delegation earns its cost
- Specialist task packets: scope evidence and authority explicitly
- Specialist synthesis: merge claims, not fluent summaries
- Specialist workflows: bound retries and review the full trace
- Project: coordinate a parcel-platform incident review
- Specialist-agent coordination decisions
Continue with: Terminal agents: keep the patch within owned files.
prompt engineering
specialist agents