abhiroopb/skills/swarm

Swarm

Fan out agents with mandatory redundancy to independently explore tasks, then challenge findings through adversarial follow-up, then consolidate through consensus.

Formula: Total agents = (Number of Tasks) × 6 (3 exploration + 3 adversarial per task)

Prerequisites

STOP and ask the user if:

• The task is simple enough that a single agent would suffice
• The task cannot be broken into independent units
• Output directory preference is unclear (default: ./swarm-output)

Phases

Phase 1: Task Breakdown

1. Analyze the task description and break into independent task units
2. Create output directory if needed
3. Create a checklist to track: each task × 2 phases (exploration, adversarial)
4. Calculate agent count: N tasks × 6 agents

Phase 2: Exploration (PARALLEL)

For each task, spawn 3 subagents in parallel:

• {task}-A, {task}-B, {task}-C

Each agent prompt:

• Assign the specific task
• Instruct to work independently and form own conclusions
• Request findings with file:line references, conclusions, and uncertainties
• Direct output to {output_dir}/{task}-{instance}.md

Wait for all Phase 2 agents to complete before proceeding.

Phase 3: Adversarial Challenge (PARALLEL)

For each task, spawn 3 adversarial subagents in parallel:

• {task}-adv-A, {task}-adv-B, {task}-adv-C

Each agent prompt:

• Point to Phase 2 outputs: {output_dir}/{task}-A.md, -B.md, -C.md
• Instruct to DISPROVE or challenge Phase 2 findings
• Request: claims confirmed, claims refuted (with counter-evidence), claims unverified, gaps identified
• Direct output to {output_dir}/{task}-adv-{instance}.md

Wait for all Phase 3 agents to complete before proceeding.

Phase 4: Consolidation

For each task, compare all 6 agent outputs and classify findings:

| Consensus Level | Meaning | |-----------------|---------| | Strong | All agents agree, adversarial agents confirm | | Weak | Exploration agrees but adversarial found gaps | | Disputed | Adversarial refuted with counter-evidence | | Uncertain | Mixed results, needs investigation |

Report to user:

• Claims that survived adversarial challenge (high confidence)
• Claims refuted with counter-evidence (likely wrong)
• Gaps identified by adversarial agents
• Cross-task patterns (findings appearing/refuted across multiple tasks)

Rules

• Never skip the adversarial phase—that's where errors get caught
• Phase 3 cannot start until Phase 2 completes (adversarial agents need exploration outputs)
• All agents in a phase launch in parallel
• Wait for all agents to complete before moving to next phase
• If an agent fails, note the failure and continue with remaining agents
• Update the checklist as each phase completes