Multi-agent Orchestration — References

Multi-agent architectures, the coordination schemes they borrow from, and what is known about why these systems fail.

Agent architectures

• AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (Wu et al., 2023), arXiv:2308.08155 — conversational multi-agent patterns and the speaker-selection problem at their core.
• MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework (Hong et al., 2023), arXiv:2308.00352 — role specialisation and structured intermediate artefacts as the coordination mechanism rather than free conversation.
• Generative Agents: Interactive Simulacra of Human Behavior (Park et al., 2023), arXiv:2304.03442 — a large population of interacting agents, and the clearest demonstration of how quickly interaction becomes unreadable.

Coordination schemes predating language models

• The Contract Net Protocol (Reid G. Smith, 1980) — task announcement, bidding, and award as a distributed assignment mechanism; the ancestor of most delegation designs. Cited by name.
• Blackboard systems (Barbara Hayes-Roth and others, 1980s) — independent knowledge sources cooperating through a shared, structured workspace rather than direct messaging. Cited by name.

Failure analysis

• Why Do Multi-Agent LLM Systems Fail? (Cemri et al., 2025) — an empirical taxonomy of multi-agent failures covering specification, inter-agent misalignment, and verification gaps. Cited by name; the exact preprint identifier is not recorded here.

Interoperation

• Agent2Agent (A2A) protocol — a published scheme for agents under separate ownership to exchange tasks and artefacts. Cited by name without a link.