Why do multi-agent AI systems fail?
The most detailed study so far, “Why Do Multi-Agent LLM Systems Fail?”, analysed more than 1,600 traces across seven frameworks and found 14 failure modes in three groups: system design issues, misalignment between agents, and weak task verification. Practitioner write-ups point to the same causes — vague delegation, missing shared context, and errors that compound over long runs. The authors conclude that simple fixes are not enough and that more structural solutions are needed.
- Published
- 22 October 2026
- Sources checked
- 22 October 2026
- From
- Kramiva
01
The three groups of failure
From the MAST taxonomy (Cemri et al.), built from expert annotation with high inter-annotator agreement (κ = 0.88).
- System design issues: poor task or role specification, steps repeated, conversation history lost, no clear stopping condition
- Inter-agent misalignment: agents ignore or withhold information, derail from the task, or proceed on wrong assumptions instead of asking
- Task verification: the run ends early, or nobody checks the result properly, or the check is wrong
02
What practitioners report
Anthropic's account of its research system describes the same problems from the inside: early versions spawned too many subagents for simple queries, and vague instructions led subagents to duplicate each other's work. Because agents are stateful and errors compound, a failure deep in a long run cannot simply be restarted; they built the ability to resume from where an error occurred.
Cognition's argument is that many of these failures are structural: when parallel agents cannot see each other's decisions, they make conflicting choices that no later step can reconcile.
03
Mitigations that the sources support
- Write explicit task descriptions for each agent: objective, output format, tools, boundaries
- Share full context, or avoid splitting work that depends on shared decisions
- Add a verification step with a clear pass condition, not just a final summary
- Persist state so a failure resumes from a checkpoint rather than from the start
- Trace every run so you can see which agent and which hand-off went wrong
04
Trade-offs
Each mitigation adds work: tracing, checkpoints and verification steps are infrastructure, and stronger verification adds model calls. The MAST authors are explicit that prompt tweaks and better role descriptions help only partly. For many tasks the cheapest fix is the one in the previous answer: use fewer agents.
Where Kramiva fits
Kramiva's AI services agree evaluation criteria before the build, measure them after it, and include production monitoring, so failures can be diagnosed from evidence rather than guessed at.
Sources
- Why Do Multi-Agent LLM Systems Fail? (Cemri et al.) — arXiv, 17 March 2025
- How we built our multi-agent research system — Anthropic, 13 June 2025
- Don't Build Multi-Agents — Cognition, 12 June 2025
Accessed 22 October 2026. Kramiva is not affiliated with these publishers.
Related questions
Tell us what you are building.
A short brief gets you a real reply from a founder within one business day — an honest read on fit, scope, and budget.