AGENT SKILL
agentseed-party
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-07-24
- Stars
- ★ 0
- Forks
- 0
03 SKILL.md 正文
AgentSeed Party Mode
Orchestrate multi-agent collaborative discussions using the generic roles defined in AGENTS.md.
When to use
Use when the user wants multiple perspectives, ideation, architecture review, technical evaluation, design review, risk review, or any multi-agent discussion.
Invocation examples
/party/party ideation/party architecture-review/party technical-evaluation/party design-review/party risk-review/party --mode session|hybrid|subagent/party --non-interactive "review the current architecture"
Modes
| Mode | Behaviour | Cost |
|---|---|---|
session |
Single model voices all personas | Lowest |
hybrid |
Rule-based + occasional LLM selection (default) | Balanced |
subagent |
Independent reasoning per persona | Highest |
Cost Optimisations
- Prefer hybrid / round-robin with preset bias.
- Use cheap model for speaker selection when needed.
- Maintain rolling summary of discussion.
- Max rounds 8–12 with early termination.
- User can force a specific speaker at any time.
Facilitator Responsibilities
- Load the relevant cast based on the preset.
- Keep agents in character and encourage productive disagreement.
- Surface trade-offs clearly.
- Produce concrete outputs (decisions, open questions, recommended change).
- At the end, offer a summary and optionally create a new change under
agentseed/changes/.
Presets
- ideation: Analyst + Product Manager + Architect
- architecture-review: Architect + Product Manager + Developer + QA (prefer subagent)
- technical-evaluation: Architect + Developer + QA
- design-review: UX + Architect + Product Manager
- risk-review: Architect + QA + Developer (adversarial)
Output Format
- Clear speaker labels
- Structured final summary:
- Key decisions
- Open questions / risks
- Recommended next action (link to change if created)