Context

Deep dive on "Language Model Teams as Distributed Systems" (arXiv:2603.12229) -- Princeton/MIT/Cambridge/NYU study applying distributed computing theory to LLM agent teams.

Key empirical findings:

• Decentralized teams: median 0.88x speedup (slower than single agent), 19 vs 4 test failures
• Centralized preassigned: median 1.36x speedup, O(n) messages
• Token cost scales with team size (rho=0.40) while speedup does not (rho=-0.07)
• Amdahl's Law confirmed (H=61.4, p<0.001): speedup ceiling = 1 / (1-p)

Action Items

• Amdahl ceiling for team sizing: Compute S_max = 1/(1-p) from task parallelizability before dispatching; use as team size ceiling in auto-topology selector. Currently only qualitative rules.
• Straggler gap metric: Add straggler_gap (slowest_agent - mean_completion) as 6th coordination metric. Decentralized-topology diagnostic.
• Token/speedup ratio alert: Add token_multiplier / latency_speedup per-run ratio; alert when >2.0 for decentralized runs. Paper: median 1.17 decentralized.
• O(n^2) message alert: Flag when decentralized team size grows past the point where O(n^2) coordination overhead dominates.
• File-locking priority: Paper shows 19 vs 4 test failures without file-level locking in decentralized dispatch. Bump priority for file-level locking when decentralized dispatch goes production.

Risks

• Paper tested homogeneous teams only; SynthOrg's role-differentiated agents mean exact numbers won't transfer -- directional findings are robust
• Model cost heterogeneity not modeled (Senior agents cost 5-20x more per token)
• O(n^2) message volume accelerates context fill, triggering more compaction (unmodeled interaction)

References

• arXiv:2603.12229
• GitHub: distributed-llm-teams
• Prior: arXiv:2512.08296 (Kim et al.) -- already integrated
• Related: research: audit engine against multi-agent failure patterns (swarm drift, microservices anti-patterns) #690 (multi-agent failure pattern audit)