unleash-the-agentsLaunch all available agents in parallel waves for open-ended hypothesis generation on problems where the correct domain is unknown. Use when facing a cross-domain problem with no clear starting point, when single-agent approaches have stalled, or when diverse perspectives are more valuable than deep expertise. Produces a ranked hypothesis set with convergen…
Independent GitHub snapshot · Automated classification · Original documentation governs
Consult all agents in parallel waves → diverse hypotheses for open-ended problems. Each agent reasons through unique domain lens — kabalist via gematria, martial-artist via conditional branching, contemplative by sitting w/ data. Convergence across independent perspectives = primary signal of merit.
Write brief any agent can understand regardless of domain. 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: Self-contained brief — agent receiving only this has all to reason.
If err: Can't articulate 5 examples|verify method → problem not ready for multi-agent. Narrow scope first.
List all agents, divide into waves of ~10. Order doesn't matter waves 1-2; 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 4 waves initially → may not need all (early stop 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 all agents allocated. advocatus-diaboli in Wave 3 (not later) → adversarial informs subsequent waves.
If err: < 20 agents → reduce 2-3 waves. Pattern works w/ as few as 10, weaker convergence signals.
Launch each wave parallel. sonnet model → cost efficiency (value from perspective diversity, not depth).
Path for ordinary interactive sessions. Per agent in wave → spawn as subagent via Agent tool (subagent_type = agent's type, e.g. kabalist, geometrist) w/ brief + domain 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 in wave simultaneous via Agent tool w/ run_in_background: true, coord via SendMessage under session's single implicit team. Wait for wave complete before next (inter-wave knowledge injection 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 w/ per-teammate task ownership + shutdown lifecycle (Task* tools themselves also work interactive — only team scoping doesn't):
TeamCreate({ description: "Wave 1: open-ended hypothesis generation" })TaskCreate per agent (brief + domain framing)Agent w/ subagent_type = agent's typeTaskUpdate w/ owner; monitor via TaskListSendMessage({ type: "shutdown_request" }) + start next w/ updated brief (Step 4)Built-in team coord — per-teammate ownership + lifecycle, status carried across waves — but team scoping only where TeamCreate surfaces. Interactive → 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 in 2-5 min. Failed|off-format noted but no block.
If err: > 50% wave fails → check brief clarity. Common: out template ambiguous|examples insufficient for non-domain agents.
After waves 1-2, extract emerging signal before next.
**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 stop guidance: Not every unleash needs all agents. Well-defined domains (codebase analysis) → convergence stabilizes 30-40 agents. Abstract|open-ended (unknown math transformations) → full roster adds value because correct domain genuinely unpredictable. Check convergence after each wave → top family count + null-model ratio plateaued → diminishing returns.
Prevents rediscovery (later waves re-deriving earlier finds) + directs later agents to edges.
Got: Later waves → more nuanced, targeted hypotheses addressing gaps in emerging consensus.
If err: No convergence after 2 waves → too unconstrained. Narrow scope or more examples.
After waves complete, gather responses → single doc. Dedupe by grouping into families:
Got: Ranked list of families w/ convergence counts, contributing agents, representative testable predictions.
If err: Every hypothesis unique (no convergence) → S/N too low. Need more examples or tighter out format.
Test top hypothesis vs null model → ensure convergence meaningful, not training data artifact.
Got: Top family significantly exceeds chance-level convergence and/or passes programmatic verify.
If err: Top fails verify → check 2nd-ranked. None passes → different approach (deeper single-expert, more data, reformulated examples).
Preferred timing: Wave 3, not post-synthesis. advocatus-diaboli in Wave 3 (alongside inter-wave injection) > standalone after all waves. Early challenge → Waves 4+ refine vs critique not pile on unchallenged consensus.
If adversarial part of Wave 3, this = final check. If not (ran all waves w/o it) → spawn advocatus-diaboli (or senior-researcher) now. Structured pass → spawn both as parallel subagents via Agent tool, coord via SendMessage vs 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: Counterarguments, edge cases, falsification experiment. Survives adversarial → ready for integration. Good adversarial sometimes partially defends consensus → design better than alts even if imperfect.
If err: Adversarial finds fatal flaw → feed critique → targeted follow-up wave (Tier 3+ iterative — 5-10 agents best for critique).
Unleash finds problems; teams solve. Convert verified families → actionable issues, assemble focused teams.
create-github-issues)subagent_type), coord via SendMessage:
teams/ matches → use itopaque-team (N shapeshifters, adaptive role) → handles unknown shapes w/o custom compadvocatus-diaboli, contemplative) → catches implementation risks tech missesGot: Each family maps → tracked issue w/ team. Unleash diagnosed; teams fix.
If err: Team comp doesn't match → reassign. Shapeshifters can research+design but lack write tools → lead applies code suggestions.
docs/investigations/lead-support-coordination-2026-09-15.md.forage-solutions — ant colony opt for solution spaces (complementary: narrower, deeper)build-coherence — bee democracy → select among competing approaches (after this skill to choose between top hypotheses)coordinate-reasoning — stigmergic coord for managing info flowcoordinate-swarm — broader swarm coord for distributed systemsexpand-awareness — open perception before narrow (complementary: individual agent prep)meditate — clear ctx noise before launch (recommended before Step 1)