The AI agent failure modes Faultmap maps before the build.
This taxonomy names the repeatable ways AI agents break: retry loops, false success, tool permission blast radius, retrieval drift, goal drift, memory corruption, multi-agent cascades, and cost runaway. Faultmap maps these classes from the goal, data, personas, and tools before the agent exists.
Failure modes
12
Clusters
8
Detection moment
Pre-build
Output
Map + tests
Failure modes, symptoms, pre-build signals, and tests.
Each failure class below is useful as a search answer, a design review checklist, and a first-test candidate.
Retry death spiral
Likely breakThe agent repeats the same failed action until cost, latency, or rate limits explode.
Control and autonomy
Repeated tool calls, rising spend, and no useful state change.
A flaky external tool sits on a required path and the goal rewards persistence.
Silent success reporting
Likely breakThe agent tells the user an action succeeded even though the tool never committed the change.
Tool and action integrity
User sees a success message, but the ticket, refund, booking, or database row is unchanged.
A write action lacks a read-after-write verification step.
Planner-executor desync
Structural riskThe planner chooses a task sequence the executor cannot actually perform with its tools or permissions.
Control and autonomy
The agent produces plausible plans, then stalls or improvises during execution.
The goal requires actions that are not represented in the available tool set.
Tool permission blast radius
Likely breakOne prompt injection or bad instruction becomes a broad system breach because the agent has more tool access than the task needs.
Tool and permission
A narrow task can trigger unrelated reads, writes, exports, or admin actions.
The agent can call high-impact tools on paths that should be read-only.
Embedding model drift
Structural riskRetrieval quality drops after an embedding or chunking change, but the agent still sounds confident.
Retrieval and data
The right document exists, but the agent retrieves a near miss or misses it entirely.
The data path relies on semantic retrieval with no fixed recall benchmark.
Small-step error compounding
Structural riskA tolerable per-step error rate becomes a high failure rate across a long agent workflow.
Multi-step reliability
The demo works on one step, but full production tasks fail unpredictably.
The workflow has many dependent steps and no checkpoint or rollback design.
Multi-agent cascade failure
Likely breakOne agent hands off partial, stale, or wrong state and the next agent compounds the mistake.
Multi-agent reliability
Specialist agents each look reasonable alone, but the crew fails as a system.
Multiple agents share work without explicit input and output contracts.
Memory corruption
Structural riskBad, stale, or irrelevant memory enters future decisions as if it were verified fact.
Memory and context
The agent remembers the wrong preference, old policy, or previous user state.
The agent can write durable memory from ambiguous or unverified context.
Goal drift
Structural riskThe agent starts with the right objective, then optimizes for a local subtask and forgets the real goal.
Memory and context
The workflow looks busy but ends far from the user's requested outcome.
The path includes long multi-turn work with changing context and no objective checkpoint.
Stale or conflicting context
Structural riskThe agent receives mutually inconsistent facts and chooses one without checking which is current.
Memory and context
The output cites old pricing, old policy, obsolete records, or contradictory sources.
The agent reads multiple systems of record that can disagree.
Tool output parsing failure
Likely breakThe tool returns a valid response in a shape the agent parses incorrectly.
Tool and action integrity
The agent uses the wrong field, misses an error flag, or treats partial output as complete.
A tool response feeds a later action without schema validation.
Cost runaway
Structural riskA workflow that looked cheap in pilot becomes expensive at production volume.
Cost and operations
Token use, tool calls, retries, and long-context turns rise faster than task volume.
The agent can loop, retrieve broadly, or escalate to expensive models on common paths.
Turn the taxonomy into a design review.
Pick the agent goal, list the data and tools it needs, then scan for the failure classes that sit on those paths. Faultmap automates that mapping and hands back the first test suite.
01
Map paths
Goal, personas, data, tools, and actions.
02
Find breaks
Match each path to known failure classes.
03
Write tests
Turn likely breaks into the first pass/fail suite.
Your agent has a few of these already.
Run a free Faultmap and see which failure modes are sitting in your design.