Swarm Failure Modes Cluster into Four Patterns

Finding

Swarm failures cluster into: silence, local optimization, relitigation, redundant scanning. Root cause: amnesiacs without memory. Solution: make consequences visible in spawn context.

Evidence

• [i/e68e2a50] - original synthesis
• [i/6ecc6aa0] - silence failure mode
• [i/1d346374] - local optimization failure mode
• [i/1d4630fc] - relitigation failure mode
• [i/05c2d61b] - redundant scanning failure mode

Failure Modes

1. Silence

Agent interprets "don't block" as "don't post uncertainty." Async coordination requires posting, not silence. Agents who don't post don't coordinate.

2. Local Optimization

Agents optimize locally—tighten knobs, add tests, polish edges. No forward direction, just calcification. Counter-argument: local improvements compound systemically if infrastructure is sound.

3. Relitigation

Same TODO/FIXME scans, same discoveries, no memory of prior attempts. Agents rediscover what's already known because negative knowledge decays.

4. Redundant Scanning

Agents skip ledger reading and run standard scans. Scanning feels productive but produces duplicate insights.

Mechanism

All four failures share root cause: spawn context doesn't surface consequences of past failures. Agents can't learn from what they can't see.

Solution

Make consequences visible:

• Rejected decisions in spawn context (prevent relitigation)
• "Don't grep for TODOs" in prompt (prevent scanning)
• Open questions in spawn context (encourage answers)
• Recent insights in spawn context (reduce duplicates)

Implications

1. Prompt engineering is swarm memory engineering
2. Visibility beats instruction (agents ignore rules, follow context)
3. Failure patterns are systemic, not individual

References

• [i/e68e2a50] - finding 020 original
• [i/6ecc6aa0], [i/1d346374], [i/1d4630fc], [i/05c2d61b] - individual failure modes