Multi-Agent Swarms Can Operate Autonomously Overnight

Finding

100 sequential agent spawns over 8+ hours with no human intervention produced 51 commits and maintained healthy system metrics.

Evidence

Test parameters (d/9b341ac4):

• 100 sequential spawns
• No human intervention
• No changes to focus directive

Results:

• 51 commits shipped
• All agents >4.9 artifacts/spawn (baseline: 2.0-8.6)
• Decision influence: 9.4% (improved from 0.6%)
• Compounding rate: 4.7% (stable)
• No system crashes or runaway loops

One concern: kitsuragi hit 5 consecutive spawns vs threshold of 3, indicating daemon prioritization could be tuned.

Mechanism

Overnight autonomy requires:

1. Task queue with clear priorities (inbox, backlog)
2. Loop detection (consecutive same-agent threshold)
3. Distributed attention (multiple agents with different focus areas)
4. Self-correction via peer review (decision rejection mechanism)

The swarm self-organized work allocation without explicit coordination.

Implications

1. Autonomous agent systems can sustain productive work for extended periods
2. Human oversight can be asynchronous rather than synchronous
3. Value of overnight autonomy = work completed while human unavailable
4. Risk: self-perpetuating activity vs. externally valuable work (separate validation needed)

Limitations

• Single test run (N=1)
• No control group (what would random task selection produce?)
• Quality of 51 commits not independently assessed
• "Valuable autonomy" vs "self-perpetuating activity" not distinguished

References

• [d/9b341ac4] - overnight test decision
• [i/935fd163] - test results
• [i/9ae725f4] - quality question raised