返回技能列表

unleash-the-agents

pjt222
更新于 Yesterday
3 次查看
17
2
17
在 GitHub 上查看
ai

关于

This skill launches multiple AI agents in parallel to generate diverse hypotheses for complex, cross-domain problems with no clear solution path. It's ideal when single-agent approaches fail or when broad perspective exploration is needed over deep specialization. The process outputs a ranked set of hypotheses, complete with convergence analysis and adversarial refinement.

快速安装

Claude Code

推荐
主要方式
npx skills add pjt222/agent-almanac -a claude-code
插件命令备选方式
/plugin add https://github.com/pjt222/agent-almanac
Git 克隆备选方式
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/unleash-the-agents

在 Claude Code 中复制并粘贴此命令以安装该技能

技能文档

Unleash the Agents

Consult all available agents in parallel waves to generate diverse hypotheses for open-ended problems. Each agent reasons through its unique domain lens — a kabalist finds patterns via gematria, a martial-artist proposes conditional branching, a contemplative notices structure by sitting with the data. Convergence across independent perspectives is the primary signal that a hypothesis has merit.

When to Use

  • Facing a cross-domain problem where the correct approach is unknown
  • A single-agent or single-domain approach has stalled or produced no signal
  • The problem benefits from genuinely diverse perspectives (not just more compute)
  • You need hypothesis generation, not execution (use teams for execution)
  • High-stakes decisions where missing a non-obvious angle carries real cost

Inputs

  • Required: Problem brief — a clear description of the problem, 5+ concrete examples, and what counts as a solution
  • Required: Verification method — how to test whether a hypothesis is correct (programmatic test, expert review, or null model comparison)
  • Optional: Agent subset — specific agents to include or exclude (default: all registered agents)
  • Optional: Wave size — number of agents per wave (default: 10)
  • Optional: Output format — structured template for agent responses (default: hypothesis + reasoning + confidence + testable prediction)

Procedure

Step 1: Prepare the Brief

Write a problem brief that any agent can understand regardless of domain expertise. Include:

  1. Problem statement: What you are trying to discover or decide (1-2 sentences)
  2. Examples: At least 5 concrete input/output examples or data points (more is better — 3 is too few for most agents to find patterns)
  3. Known constraints: What you already know, what has already been tried
  4. Success criteria: How to recognize a correct hypothesis
  5. Output template: The exact format you want responses in
## Brief: [Problem Title]

**Problem**: [1-2 sentence statement]

**Examples**:
1. [Input] → [Output]  (explain what's known)
2. [Input] → [Output]
3. [Input] → [Output]
4. [Input] → [Output]
5. [Input] → [Output]

**Already tried**: [List failed approaches to avoid rediscovery]

**Success looks like**: [Testable criterion]

**Respond with**:
- Hypothesis: [Your proposed mechanism in one sentence]
- Reasoning: [Why your domain expertise suggests this]
- Confidence: [low/medium/high]
- Testable prediction: [If my hypothesis is correct, then X should be true]

Got: A brief that is self-contained — an agent receiving only this text has everything needed to reason about the problem.

If fail: If you cannot articulate 5 examples or a verification method, the problem is not ready for multi-agent consultation. Narrow the scope first.

Step 2: Plan the Waves

List all available agents and divide them into waves of ~10. Ordering does not matter for the first 2 waves; for subsequent waves, inter-wave knowledge injection improves results.

# List all agents from registry
grep '  - id: ' agents/_registry.yml | sed 's/.*- id: //' | shuf

Assign agents to waves. Plan for 4 waves initially — you may not need all of them (see early stopping in Step 4).

WaveAgentsBrief variant
1-220 agentsStandard brief
310 agents + advocatus-diaboliBrief + emerging consensus + adversarial challenge
4+10 agents eachBrief + "X is confirmed. Focus on edge cases and failures."

Got: A wave assignment table with all agents allocated. Include advocatus-diaboli in Wave 3 (not later) so the adversarial pass informs subsequent waves.

If fail: If fewer than 20 agents are available, reduce to 2-3 waves. The pattern still works with as few as 10 agents, though convergence signals are weaker.

Step 3: Launch Waves

Launch each wave as parallel agents. Use sonnet model for cost efficiency (the value comes from perspective diversity, not individual depth).

Option A: TeamCreate (recommended for full unleash)

Use Claude Code's TeamCreate tool to set up a coordinated team with task tracking. TeamCreate is a deferred tool — fetch it first via ToolSearch("select:TeamCreate").

  1. Create the team:
    TeamCreate({ team_name: "unleash-wave-1", description: "Wave 1: open-ended hypothesis generation" })
    
  2. Create a task per agent using TaskCreate with the brief and domain-specific framing
  3. Spawn each agent as a teammate using the Agent tool with team_name: "unleash-wave-1" and subagent_type set to the agent's type (e.g., kabalist, geometrist)
  4. Assign tasks to teammates via TaskUpdate with owner
  5. Monitor progress via TaskList — teammates mark tasks completed as they finish
  6. Between waves, shut down the current team via SendMessage({ type: "shutdown_request" }) and create the next team with the updated brief (Step 4)

This gives you built-in coordination: a shared task list tracks which agents have responded, teammates can be messaged for follow-up, and the lead manages wave transitions through task assignment.

Option B: Raw Agent spawning (simpler, for smaller runs)

For each agent in the wave, spawn it with the brief and a domain-specific framing:

Use the [agent-name] agent to analyze this problem through your domain expertise.
[Paste the brief]
Think about this from your specific perspective as a [agent-description].
[For non-technical agents: add a domain-specific framing, e.g., "What patterns
does your tradition recognize in systems that exhibit this kind of threshold behavior?"]
Respond exactly in the requested format.

Launch all agents in a wave simultaneously using the Agent tool with run_in_background: true. Wait for the wave to complete before launching the next wave (to enable inter-wave knowledge injection in Step 4).

Choosing between options

TeamCreateRaw Agent
Best forTier 3 full unleash (40+ agents)Tier 2 panel (5-10 agents)
CoordinationTask list, messaging, ownershipFire-and-forget, manual collection
Inter-wave handoffTask status carries overMust track manually
OverheadHigher (team setup per wave)Lower (single tool call per agent)

Got: Each wave returns ~10 structured responses within 2-5 minutes. Agents that fail to respond or return off-format output are noted but do not block the pipeline.

If fail: If more than 50% of a wave fails, check the brief clarity. Common cause: the output template is ambiguous, or examples are insufficient for non-domain agents to reason about.

Step 4: Inject Inter-Wave Knowledge (and Evaluate Early Stopping)

After waves 1-2, extract the emerging signal before launching the next wave.

  1. Scan responses from completed waves for recurring themes
  2. Identify the most common hypothesis family (the convergence signal)
  3. Check the early stopping threshold: if the top family already exceeds 3x the null model expectation after 20 agents, you have strong signal. Plan Wave 3 as an adversarial + refinement wave and consider stopping after it
  4. Update the brief for the next wave:
**Update from prior waves**: [N] agents independently proposed [hypothesis family].
Build on this — what explains the remaining cases where this hypothesis fails?
Do NOT simply restate this finding. Extend, challenge, or refine it.

Early stopping guidance: Not every unleash needs all agents. For well-defined problem domains (e.g., codebase analysis), convergence often stabilizes at 30-40 agents. For abstract or open-ended problems (e.g., unknown mathematical transformations), the full roster adds value because the correct domain is genuinely unpredictable. Check convergence after each wave — if the top family's count and null-model ratio have plateaued, additional waves yield diminishing returns.

This prevents rediscovery (where later waves independently re-derive what earlier waves already found) and directs later agents toward the edges of the problem.

Got: Later waves produce more nuanced, targeted hypotheses that address gaps in the emerging consensus.

If fail: If no convergence appears after 2 waves, the problem may be too unconstrained. Consider narrowing the scope or providing more examples.

Step 5: Collect and Deduplicate

After all waves complete, gather all responses into a single document. Deduplicate by grouping hypotheses into families:

  1. Extract all hypothesis statements
  2. Cluster by mechanism (not by wording — "modular arithmetic mod 94" and "cyclic group over Z_94" are the same family)
  3. Count independent discoveries per family
  4. Rank by convergence: families discovered by more agents independently rank higher

Got: A ranked list of hypothesis families with convergence counts, contributing agents, and representative testable predictions.

If fail: If every hypothesis is unique (no convergence), the signal-to-noise ratio is too low. Either the problem needs more examples, or agents need a tighter output format.

Step 6: Verify Against Null Model

Test the top hypothesis against a null model to ensure the convergence is meaningful, not an artifact of shared training data.

  • Programmatic verification: If the hypothesis produces a testable formula or algorithm, run it against held-out examples
  • Null model: Estimate the probability that N agents would converge on the same hypothesis family by chance (e.g., if there are K reasonable hypothesis families, random convergence probability is ~N/K)
  • Threshold: Signal is meaningful if convergence exceeds 3x the null model expectation

Got: The top hypothesis family significantly exceeds chance-level convergence and/or passes programmatic verification.

If fail: If the top hypothesis fails verification, check the second-ranked family. If no family passes, the problem may require a different approach (deeper single-expert analysis, more data, or reformulated examples).

Step 7: Adversarial Refinement

Preferred timing: Wave 3, not post-synthesis. Including advocatus-diaboli in Wave 3 (alongside the inter-wave knowledge injection) is more effective than a standalone adversarial pass after all waves complete. Early challenge lets Waves 4+ refine against the critique rather than piling onto an unchallenged consensus.

If the adversarial pass was already part of Wave 3, this step becomes a final check. If not (e.g., you ran all waves without it), spawn advocatus-diaboli (or senior-researcher) now. For a structured pass, use TeamCreate to stand up a review team with both agents working in parallel against the consensus:

Here is the consensus hypothesis from [N] independent agents:
[Hypothesis]
[Supporting evidence and convergence stats]

Your job: find the strongest counterarguments. Where does this fail?
What alternative explanations are equally consistent with the evidence?
What experiment would definitively falsify this hypothesis?

Got: A set of counterarguments, edge cases, and a falsification experiment. If the hypothesis survives adversarial scrutiny, it is ready for integration. A good adversarial pass sometimes partially defends the consensus — finding that the design is better than alternatives even if imperfect.

If fail: If the adversarial agent finds a fatal flaw, feed the critique back into a targeted follow-up wave (Tier 3+ iterative mode — select 5-10 agents best positioned to address the specific critique).

Step 8: Hand Off to Teams

Unleash finds problems; teams solve them. Convert verified hypothesis families into actionable issues, then assemble focused teams to resolve each.

  1. Create a GitHub issue per verified hypothesis family (use the create-github-issues skill)
  2. Prioritize issues by convergence strength and impact
  3. For each issue, assemble a small team via TeamCreate:
    • If a predefined team definition in teams/ matches the problem domain, use it
    • If no fitting team exists, default to opaque-team (N shapeshifters with adaptive role assignment) — it handles unknown problem shapes without requiring a custom composition
    • Include at least one non-technical agent (e.g., advocatus-diaboli, contemplative) — they catch implementation risks that technical agents miss
    • Use REST checkpoints between phases to prevent rushing
  4. The pipeline is: unleash → triage → team-per-issue → resolve

Got: Each hypothesis family maps to a tracked issue with a team assigned. The unleash produced the diagnosis; the teams produce the fix.

If fail: If the team composition doesn't match the problem, reassign. Shapeshifter agents can research and design but lack write tools — the team lead must apply their code suggestions.

Validation

  • All available agents were consulted (or a deliberate subset was chosen with justification)
  • Responses were collected in a structured, parseable format
  • Hypotheses were deduplicated and ranked by independent convergence
  • The top hypothesis was verified against a null model or programmatic test
  • An adversarial pass challenged the consensus
  • The final hypothesis includes testable predictions and known limitations

Pitfalls

  • Too few examples in the brief: Agents need 5+ examples to find patterns. With 3 examples, most agents resort to surface-level pattern matching or template echo (repeating the brief back in different words).
  • No verification path: Without a way to test hypotheses, you cannot distinguish signal from noise. Convergence alone is necessary but not sufficient.
  • Metaphorical responses: Domain-specialist agents (mystic, shaman, kabalist) may respond with rich metaphorical reasoning that is hard to parse programmatically. Include "Express your hypothesis as a testable formula or algorithm" in the output template.
  • Rediscovery across waves: Without inter-wave knowledge injection, waves 3-7 independently rediscover what waves 1-2 already found. Always update the brief between waves.
  • Over-interpreting convergence: 43% convergence on a mechanism family sounds impressive, but check the base rate. If there are only 3 plausible mechanism families, random convergence would be ~33%.
  • Expecting single-family dominance: Abstract problems (pattern recognition, cryptography) tend to produce one dominant hypothesis family. Multi-dimensional problems (codebase analysis, system design) produce broader convergence across multiple valid families — this is expected and healthy, not a failure of the pattern.
  • Generic framing for non-technical agents: The quality of a non-technical agent's contribution depends on how the brief frames the problem in their domain language. "What does your tradition say about systems at this threshold?" produces structural insight; a generic brief produces nothing. Invest in domain-specific framing for agents outside the problem's natural domain.
  • Using this for execution: This pattern generates hypotheses, not implementations. Once you have verified hypotheses, convert them to issues and hand off to teams (Step 8). The pipeline is unleash → triage → team-per-issue.

Related Skills

  • forage-solutions — ant colony optimization for exploring solution spaces (complementary: narrower scope, deeper exploration)
  • build-coherence — bee democracy for selecting among competing approaches (use after this skill to choose between top hypotheses)
  • coordinate-reasoning — stigmergic coordination for managing information flow between agents
  • coordinate-swarm — broader swarm coordination patterns for distributed systems
  • expand-awareness — open perception before narrowing (complementary: use as individual agent preparation)
  • meditate — clear context noise before launching (recommended before Step 1)

GitHub 仓库

pjt222/agent-almanac
路径: i18n/caveman-lite/skills/unleash-the-agents
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

相关推荐技能

content-collections

Content Collections 是一个 TypeScript 优先的构建工具,可将本地 Markdown/MDX 文件转换为类型安全的数据集合。它专为构建博客、文档站和内容密集型 Vite+React 应用而设计,提供基于 Zod 的自动模式验证。该工具涵盖从 Vite 插件配置、MDX 编译到生产环境部署的完整工作流。

查看技能

polymarket

这个Claude Skill为开发者提供完整的Polymarket预测市场开发支持,涵盖API调用、交易执行和市场数据分析。关键特性包括实时WebSocket数据流,可监控实时交易、订单和市场动态。开发者可用它构建预测市场应用、实施交易策略并集成实时市场预测功能。

查看技能

creating-opencode-plugins

该Skill帮助开发者创建OpenCode插件,用于接入命令、文件、LSP等25+种事件。它提供了插件结构、事件API规范和JavaScript/TypeScript实现模式,适合需要拦截操作、扩展功能或自定义事件处理的场景。开发者可通过它快速构建响应式模块来增强OpenCode AI助手的能力。

查看技能

sglang

SGLang是一个专为LLM设计的高性能推理框架,特别适用于需要结构化输出的场景。它通过RadixAttention前缀缓存技术,在处理JSON、正则表达式、工具调用等具有重复前缀的复杂工作流时,能实现极速生成。如果你正在构建智能体或多轮对话系统,并追求远超vLLM的推理性能,SGLang是理想选择。

查看技能