Agent Exploration

Agent Exploration

Use when explicitly requested for research across independent areas with written results. Use the current harness's native explorers and configured model/effort; an explicit user override wins. Do not invoke an external runtime or create an agent registry for this workflow.

Scope and dispatch

1. Resolve the question and output path from the request or active task. Ask only for genuinely missing scope; creating the requested directory needs no additional confirmation. Default an unspecified artifact location to .audits/exploration/<timestamp>/ and state it.
2. Reuse current grounding. Scout only enough to identify independent questions; no tool-call or source-count quota. Choose as many useful slices as capacity and scope justify, up to 8; --agents is an upper bound, not a target to pad.
3. Read references/dispatch-rules.md and assets/explorer-prompt.md once when preparing dispatch. Send each explorer its question, source scope, exact output path, and the scoped-write contract once. Include assets/analysis-template.md when a structured artifact is needed; do not duplicate the contract in several embedded documents.
4. Create the output directory, then dispatch independent slices as capacity permits. Prompts need durable copies only for a long/resumable run. Every worker is read-only except for its named analysis file; no commits or unrelated edits.
5. Inspect each completed artifact and its material citations. A failed slice does not invalidate successful siblings. Continue or correct the affected worker; do not launch a fresh whole round for a missing heading or confirmation sentence.
6. Write analysis/summary.md with the answer, evidence links, disagreements, remaining gaps, and actionable conclusions. Distinguish completed and unavailable slices; never claim an unperformed investigation succeeded.

Use analysis/NN_analysis_<slug>.md for each slice. Preserve externally required schemas; otherwise keep sections as short as their evidence permits. references/checklist.md is an inspection aid, not an extra approval stage. Routine single-file lookups need no delegation or research artifact.