native-agent-deliberation ยท diff
git:20260904.09e82ca to git:20260912.7ff2e86
8 added, 45 removed. Audit A to A.
---
name: native-agent-deliberation
- description: Frame consequential or ambiguous decisions through distinct evidence-backed viewpoints, tradeoffs, and dissent. Use for brainstorming among competing strategies, architecture choices, risk review, or a decision whose assumptions should be challenged before execution.
+ description: Compare competing strategies through independent agent viewpoints when a consequential decision needs evidence, tradeoffs, and dissent.
---
- # Native Agent Deliberation
-
- ## Set The Decision
-
- 1. State one decision question.
- 2. List hard constraints, success criteria, and any real decision deadline; record "none specified" when absent.
- 3. Separate known evidence from assumptions.
- 4. Identify what would change the decision.
- 5. Keep implementation outside scope unless the user also authorizes it.
-
- ## Gather Viewpoints
-
- Load [native-subagents-first](../native-subagents-first/SKILL.md) before using native agents. Follow it as the sole lifecycle, model, budget, resume, and tool contract.
-
- Choose two to four genuinely distinct viewpoints, such as:
-
- - evidence and constraints,
- - viable options and tradeoffs,
- - dissent and failure modes,
- - reversibility and verification.
-
- Give each viewpoint a bounded question and only the evidence it needs. Ask for:
-
- - a concise recommendation,
- - supporting evidence,
- - the strongest counterargument,
- - confidence and missing facts,
- - a condition that would reverse the recommendation.
+ # Native agent deliberation
- Use the tool's structured role controls when needed. Do not put role or `agent_type` markers in prompt text.
+ Frame one decision with its constraints, success criteria, known evidence, assumptions, and any real deadline. Identify what evidence could change the choice.
- ## Synthesize
+ Before using agents, load [native-subagents-first](../native-subagents-first/SKILL.md) for ownership, model selection, budgets, and lifecycle rules. Select distinct viewpoints that can change the decision, such as feasibility and failure risk. Do not create a fixed number of roles merely to fill a panel.
- 1. Verify material claims against source evidence.
- 2. Compare options against the same criteria.
- 3. Preserve disagreements instead of manufacturing consensus.
- 4. Distinguish evidence, inference, and preference.
- 5. Recommend one path, defer, or ask for one blocking fact.
- 6. Name the strongest dissent and what would change the decision.
- 7. After deciding, continue work already authorized by the user. Otherwise give the next concrete action without starting unrequested implementation.
+ Give each agent a bounded question and the evidence it needs. Ask for a recommendation, supporting facts, strongest counterargument, uncertainty, and a condition that would reverse the recommendation. Use structured tool controls for roles, not role markers in prompt text.
- ## Return
+ Verify material claims and compare options against the same criteria. Preserve real disagreement; distinguish evidence from inference and preference. Recommend a path or identify the specific missing fact that blocks the choice.
- - Decision and recommendation.
- - Evidence that drove it.
- - Strongest tradeoff or dissent.
- - Uncertainty and reversal condition.
- - Next action.
+ Return the recommendation, decisive evidence, strongest tradeoff or dissent, uncertainty, and next action. Continue implementation when already authorized. A request only for deliberation does not authorize implementation.
- Do not use this skill for generic subagent execution, routine implementation fanout, or validation.
+ Use execution or validation skills for ordinary implementation and testing rather than turning those tasks into a debate.