CH 04 · Why Multi-Agent Systems Fail: The MAST Taxonomy · 04/08
Read as one pageThe three categories of failure
MAST sorts its 14 failure modes by the stage of the agent lifecycle where they originate: before execution, during execution, and after execution.
1. Specification issues (about 41.8 percent of failures)
These come from a flawed setup: poor prompt design, missing role constraints, or no clear stopping criteria. Representative modes include:
- Disobeying the task specification.
- Repeating steps that were already completed.
- Losing track of the conversation history.
- Failing to recognize when the task is actually done.
This is the largest category, which is a striking result on its own: the single biggest source of failure is how the system was specified and set up, before any agent even starts talking to another.
2. Inter-agent misalignment (about 36.9 percent)
These happen during execution, from miscommunication, conflicting assumptions, or context that never gets passed along. Examples include:
- Ignoring what other agents said.
- Proceeding without asking for clarification.
- Resetting conversations unexpectedly.
- A mismatch between what an agent reasons and what it actually does.
This category is essentially the empirical confirmation of Cognition's Principle 2 from Chapter 3: conflicting hidden decisions and unshared context are not just a theoretical worry, they account for over a third of observed failures.
3. Task verification failures (about 21.3 percent)
These come from weak quality control at the end:
- Ending the task too early.
- Skipping validation entirely.
- Accepting an incorrect solution because the check was shallow.
The article highlights a recurring pattern here: many systems do include a verifier agent, but its checks are superficial. Code is accepted just because it compiles. A program is assumed correct because its comments look consistent. That is not real verification.
An important note: many traces contain several failure modes at once, which is why the paper argues for structured diagnosis rather than ad hoc inspection.
Finished this idea?
Your learning progress stays in this browser.