Agent Expert Panel
Decide whether a task needs a coordinated expert team, then design the smallest reliable workflow. A panel is an accountable workflow, not a list of personas.
Contract
Design or audit by default; never launch agents or invent evidence. Operationalize only when explicitly requested and supported by the host. Prefer one agent unless a panel adds enough quality or risk reduction to justify its cost.
Procedure
- Intake: capture outcome, acceptance evidence, inputs, capabilities, and material budget, privacy, approval, or rollback constraints.
- Fit: form a panel only when separable work, distinct expertise, checking, or integration value outweighs operating cost. Never decide by task count alone.
- Path: choose
direct, compact-panel, or full-panel; reserve full for a reusable contract or implementation handoff.
- Ownership: the steward owns the final outcome and trace. Each specialist owns one artifact with focused inputs, limits, a check, and a next owner.
- Preflight: record tools, data sensitivity, budget, approval, and fallback. Mark each packet
match, mismatch, or clarify; assume nothing.
- Flow: default to sequential work; parallelize only independent packets with merge criteria. Handoffs need an owner, artifact, check, stop condition, and next owner. Gate publication, spend, sensitive transfer, and irreversible work.
- Failure: stop on missing capability, replan repeated failure, return ambiguous work for evidence, and escalate irreversible or value-based conflicts.
Output
- Direct: decision, reason, next action, and when to reconsider.
- Compact (default): assumptions, members, phases, handoffs, gates, checks, risks, fallback, and next action.
- Full: read
references/panel-specification.md and apply its detailed contracts.
Do not silently execute a design request or emit a full specification for a small task.
Resources
Use evals/ for behavior changes, scripts/validate_package.py for packaging, and reports/output_quality_scorecard.md for claim boundaries.