Multi-Agent Coordination Failures¶

Multi-agent systems introduce coordination failure modes that are distinct from single-agent LLM failures. These emerge from the interaction of multiple bounded agents operating in a shared space, each with limited context and potentially different information states.

Failure Modes¶

Flooding¶

Multiple agents respond to the same event simultaneously, producing output that exceeds participants' processing capacity. In a system with eight agents, a single significant event can trigger eight near-simultaneous assessments that overwhelm both human operators and other agents trying to maintain coherent situation-awareness-in-human-ai-teams.

Mitigation: delivery-modes — flow gating prevents simultaneous arrival; priority ordering ensures the most safety-relevant outputs reach the operator first.

Staleness¶

An agent's context snapshot may not reflect the current shared state by the time its response arrives. During dynamic situations where conditions change on a timescale of seconds, an agent reasoning about conditions from 30 seconds ago may produce recommendations that conflict with the current state. Actions taken on stale context can conflict with other agents' more recent actions.

Mitigation: Staleness checks before action; timestamps on all data in context; shorter reasoning chains during fast transients.

Context Divergence¶

Agents build their contexts differently depending on what they have attended to, what their memory-architectures retrieves, and how much history fits in their context-windows. Two agents may hold significantly different models of the current situation — not because they reason differently, but because they perceive different histories.

Context window arithmetic: In a room with eight agents generating 300 tokens per response at 45-second intervals, combined output is roughly 3,200 tokens per minute. A 128K-token context window fills in approximately 38 minutes. In a four-hour session, later agents work from compressed, partial views. SA divergence between early and late agents is guaranteed by arithmetic.

Cascading Tool Use¶

A tool-calling extends an agent's turn while the environment evolves. Other agents may respond to the partial state in ways that conflict with the eventually-returned result.

Epistemic Asymmetry from Threading¶

An agent engaged in a sub-thread has a narrower view than an agent attending to the main room. They hold different situational pictures simultaneously.

Heartbeat-Delivery Interaction¶

A heartbeat-triggered agent enters the room not in response to any message. Standard delivery modes do not cleanly handle this — the agent must either pre-empt the current flow (potentially interrupting important exchanges) or queue (potentially delaying time-sensitive monitoring).

Conversational Protocol Failures¶

Li et al. (2023) document four failure modes specific to agents coordinating through natural language conversation:

Role flipping — An agent assigned to follow instructions begins issuing them instead, reversing the intended power dynamic. The other agent may comply, producing a conversation that looks functional but has inverted authority. Mitigation requires explicit role-identity enforcement in system prompts.
Instruction repetition — An agent echoes instructions back without adding substance. The conversation loops without progress but may not trigger timeout conditions because messages are being exchanged.
Flake replies — Agents produce superficially responsive outputs ("I will do something") without concrete content. The conversation appears to advance but no real work is done. This is a conversational analogue of output-vacuity.
Infinite message loops — Both agents enter a cycle of pleasantries, gratitude, or goodbyes without recognising the task is complete or stuck. Explicit termination tokens and message count limits are required to break these loops.

These conversational failures are distinct from the coordination failures above: they arise not from concurrent operation or context limitations, but from the inability of language model agents to maintain stable conversational protocols over extended exchanges. They are most prevalent in Pattern 4 (Role-Based Crew) and Pattern 5 (Group Chat) architectures where agents coordinate through free-form dialogue.

Coordination Paralysis in Reasoning Models¶

Riedl (2025) documents a failure mode specific to reasoning models (observed in QWEN3): when prompted to reason about other agents' likely actions (theory of mind) under noisy group feedback, agents enter infinite chain-of-thought loops attempting to reconcile their local strategy with uncertain information about group dynamics. The reasoning process itself becomes a trap — the more the agent reasons about coordination, the less able it is to act. This "paralysis under coordination ambiguity" is distinct from the conversational protocol failures above; it arises from the depth of reasoning rather than from conversational dynamics.

Infrastructure Failures¶

The coordination infrastructure itself (delivery modes, threading, governance-gates) is software that can fail, producing system-level errors distinct from individual agent errors.

The MAST Taxonomy: Empirical Failure Rates¶

Cemri et al. (2025) provide the first empirically grounded taxonomy from 1,642 annotated failure traces across 7 MAS frameworks. Failure rates range from 41% to 86.7% across systems. The 14 failure modes cluster into three categories:

System Design Issues (44.2%) — the largest category, representing failures from architectural choices rather than model limitations: - Step repetition (15.7%), unaware of termination conditions (12.4%), disobey task specification (11.8%), loss of conversation history (2.8%), disobey role specification (1.5%)

Inter-Agent Misalignment (32.5%) — breakdowns in agent coordination: - Reasoning-action mismatch (13.2%), task derailment (7.4%), fail to ask for clarification (6.8%), conversation reset (2.2%), ignored other agent's input (1.9%), information withholding (0.85%)

Task Verification (23.5%) — quality control failures: - Incorrect verification (9.1%), no or incomplete verification (8.2%), premature termination (6.2%)

Key insight: 44.2% of failures are design issues fixable through better architecture — not model limitations. A simple workflow fix (ensuring the CEO agent had final authority) improved ChatDev's success rate by +9.4% with the same underlying model. This validates that multi-agent-taxonomy pattern selection determines system properties more than model capability.

Dependency Graph Defects¶

Gao et al. (2025) model multi-agent execution as a directed dependency graph and identify three structural defect categories that explain why MAS can underperform single-agent systems:

Node-level defect: Performance is bottlenecked by the critical agent handling the hardest subtask. Task decomposition across multiple agents does not help if the core difficulty exceeds any single LLM's capability. ~80% of test cases are ties (both MAS and SAS succeed or both fail), confirming this is the dominant factor.
Edge-level defect: Downstream agents are overwhelmed by upstream inputs. In high-in-degree nodes of the execution graph, excessive information from multiple sources causes overthinking and incorrect outputs — analogous to the flooding failure mode above, but at the individual agent level rather than the human operator level.
Path-level defect: Errors propagate through chains of agent interactions. When intermediate outputs are summarised or filtered, crucial context is lost irreversibly. Unlike single-agent systems where the full history remains in context, MAS information loss cascades downstream.

These defects are structural: they arise from the graph topology of multi-agent coordination, not from model capability. They explain the empirical finding that MAS advantage over single-agent diminishes from 10–16% to 0.8–3% as frontier LLMs improve, while MAS costs remain 4–220x higher.

Coordination Failure Modes by Pattern¶

Pattern	Primary Failure	Mechanism
0: Role Simulation	Correlated errors	All roles share one context and model
1: Autonomous Loop	Goal drift	No external constraint on task queue
2: Sequential Pipeline	Cascading errors	Error in stage N propagates downstream
3: Orchestrator-Worker	Single point of failure	Orchestrator overload or poor decomposition
5: Group Chat	Speaker selection loops	Repetitive turns without progress
6: Handoff/Transfer	Context loss at handoff	Incomplete state transfer
9: Shared Room	Flooding, staleness, divergence	Concurrent operation with bounded context

Degradation and Recovery¶

Multi-agent systems require defined degraded-mode operation:

Agent failure detection: silent failure (no output), explicit failure (error signal), degraded failure (lower quality output). Silent and degraded failures are hardest to detect.
Graceful degradation: loss of different agent roles produces different residual capabilities. Loss of the adversarial agent removes independent challenge; loss of the synthesiser forces manual integration.
State reconstruction: when a failed agent restarts, it has empty context. Pre-computed state summaries or periodic checkpoints reduce the SA gap.
Fallback modes: reduced agent count → reduced model diversity → single-agent → no-AI operation. Each should be a defined system state with its own display mode.

Relevance to Safety-Critical Systems¶

These failures are emergent, not individual. They arise from the interaction of multiple agents, not from any single agent's malfunction. Testing individual agents in isolation will not reveal coordination failures.
Context window arithmetic makes divergence inevitable. In any session longer than ~40 minutes with active multi-agent operation, agents will have different information states. System design must accommodate this, not assume it away.
Cross-domain examples. In aviation, multiple advisory systems providing conflicting guidance during a critical phase of flight. In medical settings, multiple AI diagnostic tools accessing different subsets of patient data and producing divergent assessments. In oil and gas, multiple monitoring systems providing asynchronous readings during a well control event.
The coordination infrastructure is itself a failure surface. Delivery modes, threading, and governance gates must be tested as components, not just the agents they coordinate.