unleash-the-agentsLaunch all available agents in parallel waves for open-ended hypothesis generation on problems where correct domain unknown. Use when face cross-domain problem with no clear starting point, when single-agent approaches stalled, or when diverse perspectives more valuable than deep expertise. Produces ranked hypothesis set with convergence analysis, adversari…
Independent GitHub snapshot · Automated classification · Original documentation governs
Consult all available agents in parallel waves. Generate diverse hypotheses for open-ended problems. Each agent reasons through unique domain lens — kabalist finds patterns via gematria, martial-artist proposes conditional branching, contemplative notices structure by sitting with data. Convergence across independent perspectives = primary signal hypothesis has merit.
Write problem brief any agent can understand regardless of domain expertise. Include:
## 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: Brief self-contained — agent receiving only this text has all needed to reason about problem.
If fail: Cannot articulate 5 examples or verification method? Problem not ready for multi-agent consultation. Narrow scope first.
List all available agents. Divide into waves of ~10. Ordering matters not for 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 initial. May not need all (see early stopping in Step 4).
| Wave | Agents | Brief variant |
|---|---|---|
| 1-2 | 20 agents | Standard brief |
| 3 | 10 agents + advocatus-diaboli | Brief + emerging consensus + adversarial challenge |
| 4+ | 10 agents each | Brief + "X is confirmed. Focus on edge cases and failures." |
Got: Wave assignment table with all agents allocated. Include advocatus-diaboli in Wave 3 (not later) so adversarial pass informs subsequent waves.
If fail: Fewer than 20 agents available? Reduce to 2-3 waves. Pattern still works with as few as 10 agents. Convergence signals weaker.
Launch each wave as parallel agents. Use sonnet model for cost efficiency (value comes from perspective diversity, not individual depth).
Path for ordinary interactive sessions. For each agent in wave, spawn as subagent via Agent tool (subagent_type set to agent's type, e.g. kabalist, geometrist) with brief and 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 wave simultaneous via Agent tool with run_in_background: true. Coordinate with SendMessage under session's single implicit team. Wait for wave to complete before launching next (enable inter-wave knowledge injection in Step 4).
TeamCreate deprecated, gated out of ordinary interactive sessions — ToolSearch("select:TeamCreate") returns nothing there, team_name ignored (session has single implicit team). Where it does surface (FleetView / cloud), adds named team object with per-teammate task ownership and shutdown lifecycle (Task* tools themselves also work in interactive sessions — only team scoping not):
TeamCreate({ description: "Wave 1: open-ended hypothesis generation" })TaskCreate (brief + domain-specific framing)Agent tool with subagent_type set to agent's typeTaskUpdate with owner; monitor with TaskListSendMessage({ type: "shutdown_request" }), start next with updated brief (Step 4)Built-in team coordination — per-teammate ownership and lifecycle, status carried across waves — but team scoping only exists where TeamCreate surfaces. In interactive sessions use Option A.
| Agent tool (Option A) | TeamCreate (Option B) | |
|---|---|---|
| Availability | Every session (primary) | FleetView / cloud only (gated) |
| Best for | All interactive unleashes | Cloud runs wanting shared task list |
| Coordination | SendMessage, manual collection | Task list, messaging, ownership |
| Inter-wave handoff | Track via brief update | Task status carries over |
Got: Each wave returns ~10 structured responses within 2-5 minutes. Agents that fail to respond or return off-format output noted but don't block pipeline.
If fail: More than 50% of wave fails? Check brief clarity. Common cause: output template ambiguous, or examples insufficient for non-domain agents to reason about.
After waves 1-2, extract emerging signal before launching 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 (codebase analysis), convergence often stabilizes at 30-40 agents. For abstract or open-ended problems (unknown mathematical transformations), full roster adds value because correct domain genuinely unpredictable. Check convergence after each wave — if top family's count and null-model ratio plateaued, additional waves yield diminishing returns.
Prevents rediscovery (where later waves independently re-derive what earlier waves already found). Directs later agents toward edges of problem.
Got: Later waves produce more nuanced, targeted hypotheses addressing gaps in emerging consensus.
If fail: No convergence after 2 waves? Problem may be too unconstrained. Consider narrowing scope or providing more examples.
After all waves complete, gather all responses into single document. Deduplicate by grouping hypotheses into families:
Got: Ranked list of hypothesis families with convergence counts, contributing agents, representative testable predictions.
If fail: Every hypothesis unique (no convergence)? Signal-to-noise ratio too low. Either problem needs more examples, or agents need tighter output format.
Test top hypothesis against null model. Ensure convergence meaningful, not artifact of shared training data.
Got: Top hypothesis family significantly exceeds chance-level convergence and/or passes programmatic verification.
If fail: Top hypothesis fails verification? Check second-ranked family. No family passes? Problem may need different approach (deeper single-expert analysis, more data, reformulated examples).
Preferred timing: Wave 3, not post-synthesis. Including advocatus-diaboli in Wave 3 (alongside inter-wave knowledge injection) more effective than standalone adversarial pass after all waves complete. Early challenge lets Waves 4+ refine against critique rather than piling onto unchallenged consensus.
Adversarial pass already part of Wave 3? This step becomes final check. Not? (e.g., ran all waves without it) — spawn advocatus-diaboli (or senior-researcher) now. For structured pass, spawn both agents as parallel subagents via Agent tool, coordinate with SendMessage against 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: Set of counterarguments, edge cases, falsification experiment. Hypothesis survives adversarial scrutiny? Ready for integration. Good adversarial pass sometimes partially defends consensus — finds design better than alternatives even if imperfect.
If fail: Adversarial agent finds fatal flaw? Feed critique back into targeted follow-up wave (Tier 3+ iterative mode — select 5-10 agents best positioned to address specific critique).
Unleash finds problems. Teams solve them. Convert verified hypothesis families into actionable issues. Assemble focused teams to resolve each.
create-github-issues skill)subagent_type), coordinate with SendMessage:
teams/ matches problem domain? Use itopaque-team (N shapeshifters with adaptive role assignment) — handles unknown problem shapes without requiring custom compositionadvocatus-diaboli, contemplative) — they catch implementation risks technical agents missGot: Each hypothesis family maps to tracked issue with team assigned. Unleash produced diagnosis. Teams produce fix.
If fail: Team composition doesn't match problem? Reassign. Shapeshifter agents can research and design but lack write tools — team lead must apply their code suggestions.
docs/investigations/lead-support-coordination-2026-09-15.md.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 agentscoordinate-swarm — broader swarm coordination patterns for distributed systemsexpand-awareness — open perception before narrowing (complementary: use as individual agent preparation)meditate — clear context noise before launching (recommended before Step 1)