MAST: Multi-Agent System Failure Taxonomy
Cemri et al. (UC Berkeley + collaborators) · varXiv 2503.13657v3 (2025-10-26) · first published 2025-03-17
Scope
Failure modes of multi-agent LLM systems: 14 failure modes in 3 categories, derived from expert annotation of 150 traces (kappa = 0.88) and applied to MAST-Data (1,600+ annotated traces across 7 multi-agent frameworks).
Why it's ingested
MAST is the most-cited 2025 multi-agent failure taxonomy and is a primary ingestion target for SAICA-KG's crosswalk contribution. Its MAS focus complements SAICA-KG's single-agent coding-agent scope — crosswalks to FailureMode tags surface which modes generalize across agent topologies.
Categories & crosswalks
Each row is one category of this taxonomy with the SAICA failure modes mapped to it and how close each mapping is: “exact” means the same failure, “partial” an overlap, “related” a neighbouring category. 14 mappings cover 10 of the 17 categories; the 7 with no mapping are listed under them. The category ids and labels were last compared with the source on 2026-11-21.
Mapping drafted by an AI model from the source text; not yet reviewed by a person.
-
FM-1.1Disobey task specificationFailure to adhere to the specified constraints or requirements of a given task, leading to suboptimal or incorrect outcomes.
SAICA-KG crosswalks: -
FM-1.2Disobey role specificationFailure to adhere to the defined responsibilities and constraints of an assigned role, potentially leading to an agent behaving like another.
SAICA-KG crosswalks: -
FM-1.3Step repetitionUnnecessary reiteration of previously completed steps in a process, potentially causing delays or errors in task completion.
SAICA-KG crosswalks: -
FM-1.4Loss of conversation historyUnexpected context truncation, disregarding recent interaction history and reverting to an antecedent conversational state.
SAICA-KG crosswalks: -
FM-1.5Unaware of termination conditionsLack of recognition or understanding of the criteria that should trigger the termination of the agents' interaction, potentially leading to unnecessary continuation.
SAICA-KG crosswalks: -
FM-2.1Conversation resetUnexpected or unwarranted restarting of a dialogue, potentially losing context and progress made in the interaction.
SAICA-KG crosswalks: -
FM-2.3Task derailmentDeviation from the intended objective or focus of a given task, potentially resulting in irrelevant or unproductive actions.
SAICA-KG crosswalks: -
FM-3.1Premature terminationEnding a dialogue, interaction or task before all necessary information has been exchanged or objectives have been met, potentially resulting in incomplete or incorrect outcomes.
SAICA-KG crosswalks: -
FM-3.2No or incomplete verification(partial) omission of proper checking or confirmation of task outcomes or system outputs, potentially allowing errors or inconsistencies to propagate undetected.
SAICA-KG crosswalks: -
FM-3.3Incorrect verificationFailure to adequately validate or cross-check crucial information or decisions during the iterations, potentially leading to errors or vulnerabilities in the system.
SAICA-KG crosswalks:
Categories with no SAICA failure mode
These 7 categories have no mapping. The reason is the one the crosswalk file states; where it states none, that is said.
-
FC1System Design IssuesFailures originate from system design decisions, and poor or ambiguous prompt specifications.
Why it is not mapped: The crosswalk file gives no reason.
-
FC2Inter-Agent MisalignmentFailures arise from a breakdown in critical information flow from inter-agent interaction and coordination during execution.
Why it is not mapped: The crosswalk file gives no reason.
-
FC3Task VerificationFailures involve inadequate verification processes that fail to detect or correct errors, or premature termination of tasks.
Why it is not mapped: The crosswalk file gives no reason.
-
FM-2.2Fail to ask for clarificationInability to request additional information when faced with unclear or incomplete data, potentially resulting in incorrect actions.
Why it is not mapped: Fail to ask for clarification: SAICA has no mode for not asking or not escalating.
-
FM-2.4Information withholdingFailure to share or communicate important data or insights that an agent possess and could impact decision-making of other agents if shared.
Why it is not mapped: Information withholding: between agents only.
-
FM-2.5Ignored other agent's inputDisregarding or failing to adequately consider input or recommendations provided by other agents in the system, potentially leading to suboptimal decisions or missed opportunities for collaboration.
Why it is not mapped: Ignored other agent's input: between agents only.
-
FM-2.6Reasoning-action mismatchDiscrepancy between the logical reasoning process and the actual actions taken by the agent, potentially resulting in unexpected or undesired behaviors.
Why it is not mapped: Reasoning-action mismatch: no SAICA mode is defined by a gap between what the agent reasons and what it does.
Notes
Checked 2026-11-21 against arXiv 2503.13657v3 (2025-10-26, title "Why Do Multi-Agent LLM Systems Fail?"): category names and definitions from Section 4, failure-mode names and definitions from Appendix A. Earlier versions use the same FM-1.1 to FM-3.3 ids and mode names, but v1 names the categories "Specification and System Design Failures" (FC1) and "Task Verification and Termination" (FC3). Earlier SAICA-KG releases carried paraphrased seed labels (MAST-FM-01 to MAST-FM-14) that did not match the paper; they were replaced by the paper's ids on 2026-11-21 and all crosswalks were re-judged against these definitions.