Multi-Agent Systems Have a Distributed Systems Problem

Christopher Meiklejohn — a decade-deep CRDT and distributed-systems researcher (Lasp, Partisan, Filibuster, Basho) — argues that multi-agent LLM systems reproduce, by construction rather than by accident, every classic failure category of distributed systems: lost updates, stale reads, crash recovery, causal ordering, partitions, and Byzantine faults. His evidence is hands-on: two Claude Code instances in different worktrees both wrote migration 267 and one silently clobbered the other.

The Argument

The essay opens with a story that is also a diagnosis. Two agents, two worktrees, one migration filename, different schemas, silent overwrite. Meiklejohn's reaction — laughter — is the tell of someone recognising an old enemy in new clothes: a lost update, the problem CRDTs were invented for, playing out in a directory of SQL files.

His survey of the multi-agent literature lands the punch:

Shared mutable state with no formal concurrency control. No fault model. No reasoning about what happens when agents disagree.

ChatDev has "communicative dehallucination" (a role-reversal clarification heuristic) but no causal ordering across chat chains. MetaGPT's shared message pool is real progress — publish-subscribe with role-profile subscriptions — but pub-sub is not concurrency control: it tells you what other agents produced, not whether you're reading a stale version. AutoGen has no persistent shared state at all. Three of the field's flagship frameworks, one shared blind spot.

The centrepiece is a five-item taxonomy of inevitable failures:

• Conflicts and stale reads — two agents edit the same file; one's work vanishes. Or an agent picks a bug from the tracker that was fixed ten minutes ago.
• Failure and recovery — the crash-recovery model, but the replicas hit context limits and hallucinate instead of merely segfaulting.
• Ordering without a clock — a triage agent assigns a bug to A while B, who saw the report first, is already fixing it. Lamport 1978 and vector clocks solve this; "the multi-agent world hasn't noticed yet."
• Partition tolerance — API timeouts and rate limits split agents into diverging groups whose states must merge losslessly on reconnect.
• Byzantine faults — an agent hallucinates a plausible fix, passes its own tests, ships it; downstream agents build on top. "Every agent is a potential Byzantine actor every time it responds."

The honest coda matters: he doesn't claim the distributed-systems toolkit transfers cleanly. LLM agents aren't database replicas — they hallucinate, lose context, and make confident decisions on incomplete information. Whether CRDTs, version vectors, and fault injection carry over is, to him, "the most interesting question I've encountered in a long time."

Key Themes

#concept #pattern — coordination as the missing layer beneath agent frameworks

Analysis

The strongest thing about this essay is provenance: most "multi-agent systems are just distributed systems" takes come from people who arrived at agents first. Meiklejohn arrived from the other direction — Riak, SyncFree, a 1,000-node CRDT deployment, a 10-year influence award — and his authority is earned, not asserted. That makes the essay's restraint credible: he is not saying "use CRDTs," he is saying the problem shape is identical and the solution space is unexplored.

Two complications worth pushing on. First, the Byzantine framing is doing heavy lifting: distributed systems handle Byzantine faults with quorum thresholds (3f+1 replicas), but an agent's hallucinations aren't adversarial, they're stochastic — more like soft errors than traitors, and possibly better served by the verification-heavy patterns the agent world has already developed than by Byzantine consensus. Second, and related: the essay's own caveat undercuts the taxonomy slightly. If agents differ from replicas in the exact dimensions where the hardest distributed-systems techniques bite (Byzantine behaviour, partial observability), then the transferable fraction may be the 1970s–80s core (ordering, versioning, recovery) rather than the hardest problems. That's still a big deal — stale-read detection alone would have saved migration 267.

It's also a quiet indictment of the field's literature: three well-cited frameworks, and the coordination layer underneath is message passing. Fifty years of research, none of it imported.

Relation to the Wiki

This source sharpens Agent Swarm Model Economics, which catalogued five coordination failure modes at scale empirically — Meiklejohn supplies the formal vocabulary (lost updates, causal ordering, Byzantine actors) that explains why Cursor hit those failures. It complicates Deterministic Spine, Agentic Leaves: that piece prescribes deterministic components owning state as the fix, while Meiklejohn argues the deeper gap is formal concurrency control on shared state — compatible, but his framing is more foundational. It also gives Stash — Conflict-Free Folder Sync an unexpected dignification: a practical CRDT-adjacent merge engine for exactly the kind of shared-directory state (agent memory, skills files) that produces Meiklejohn's lost updates.

Sources: raw/multi-agent-systems-have-a-distributed-systems-problem-html, Multi-Agent Systems Have a Distributed Systems Problem Last updated: 2026-09-25