Multi-Agent Orchestration Patterns: A Map

Tone: clean-academic, survey register — a map to sit next to the deeper theoretical and applied posts, not a replacement for either.

Original. “Composition does not preserve reliability… Multi-agent systems are often less reliable than the single loop they replaced, and the failure is located in the seams — task specification and result fusion — not in the workers.” — from the agent-loop field guide, Sub-problem 4.

Essence. The field guide proved why naive multi-agent composition can hurt more than it helps (reliability compounds as roughly $p^n$ across dependent steps, before you even add the orchestrator’s own decomposition/fusion error). What it didn’t do — deliberately, it’s a theory post — is catalogue the actual patterns people use once they compose loops anyway. This post is that catalogue: four real patterns, compared on the one thing the reliability math says matters — where the seams are, and what each pattern does to keep them from being the weak point.

0. The seam is the point

Every multi-agent pattern below is, underneath its name, a different answer to one question: when you split a task across multiple LLM calls, what exactly gets passed across the seam? Full reasoning traces? Final answers only? A shared draft everyone edits? The answer determines both the pattern’s failure mode and its cost.

1. Four real patterns

• Orchestrator–Workers — a controller decomposes a task and dispatches sub-tasks to specialized workers, then fuses their results. The seam carries task specifications going out and results coming back; workers don’t see each other. This is the pattern named (and formally cited only informally, via Anthropic’s own engineering writeup) in the gap post and worked out concretely in the financial-LLM post.
• Multi-Agent Debate (Du et al., 2024) (2023, ICML 2024) — multiple instances of a model propose answers, see each other’s full reasoning, and revise over several rounds toward a common answer. The seam carries complete reasoning traces, not just conclusions — the mechanism the paper credits for reducing hallucination and improving factuality.
• Mixture-of-Agents (Wang et al., 2024) (2024) — a layered architecture: each layer of agents takes every prior layer’s outputs as auxiliary context and produces a refined response, repeated for several layers. The seam carries full prior-layer outputs forward, never backward — no debate, no revision of earlier layers, only synthesis going up.
• Reflexion (Shinn et al., 2023) (2023) — technically a single agent, not multiple, but the same seam question applies across time instead of across agents: the agent critiques its own attempt in natural language, stores that verbal feedback, and retries. Included because it’s the degenerate case (n=1) that the multi-agent patterns generalize — worth knowing what composition adds over self-critique alone.

2. Comparison table

3. What the field guide’s math predicts for each

Going back to the $p^n$ reliability-composition result: Orchestrator–Workers is the pattern most exposed to it — a genuine dependency chain (decompose → delegate → fuse), so the compounding failure the field guide warns about is a real risk, not a theoretical one. Debate and Mixture-of-Agents are structurally different: they’re not a dependency chain toward one answer but multiple independent-ish attempts reconciled at the end, which is closer to an ensemble than a pipeline — a different reliability story the $p^n$ formula doesn’t directly describe. This is a genuine open question this survey doesn’t resolve: none of the four papers analyzes its own reliability composition the way the field guide’s theory post does for the general case.

4. Open questions

• Does Multi-Agent Debate’s “expose disagreement” mechanism still help once all debating agents are instances of the same model (shared blind spots), or does it need genuine model heterogeneity to work?
• Mixture-of-Agents never lets a later layer correct an earlier one’s mistake it didn’t propagate forward cleanly — is that a real limitation, or does synthesis-only composition avoid a different failure mode (infinite revision loops) that debate is exposed to?
• The kmux-observer post’s actual question — when should a human’s attention get pulled into any of these four patterns, not just when should the system escalate to a bigger model — isn’t answered by any of the four papers here. That’s still open.