Most AI agent workflows fail in a small set of predictable ways: a tool call returns an error payload wrapped in HTTP 200, a loop never reaches its exit condition, context grows until the model quietly degrades, and nobody can tell which step produced the bad output. Reliability does not come from a smarter model. It comes from mapping each failure mode to a control and instrumenting the signal that proves the control works.
This guide is that map. Every failure mode below pairs with one concrete control and one measurable signal, drawn from published guidance by Anthropic, Google Cloud, and the Chroma context-rot research.
Failure mode 1: silent tool failures
An agent calls a tool, the transport succeeds, and the payload says the operation failed — or returns empty, stale, or malformed data. If your workflow treats HTTP 200 as success, the agent proceeds on garbage.
Control: validate the content of every tool result against an expected schema, not the transport status. Treat empty or ambiguous output as failure. Anthropic's guidance on building effective agents makes the same point: agent systems need explicit error surfaces because models cannot recover from errors they cannot see.
Signal: per-tool error rate and empty-result rate in your logs. If a tool's silent-failure rate is invisible, the control does not exist.
Failure mode 2: loops that never terminate
An agent retries, replans, or calls the same tool in a cycle because no termination condition is ever satisfied. Google Cloud's agentic design guidance calls out the loop pattern's primary trade-off: an infinite loop when the exit condition is incorrectly defined.
Control: hard iteration caps plus an exit condition the workflow itself can evaluate, plus time, token, and cost budgets. The cap is the circuit breaker; the evaluable condition is the intended exit.
Signal: p95 iterations per task and cost per task. A loop problem shows up in the distribution tail before it shows up as an outage.
Failure mode 3: context rot
Chroma's July 2025 report evaluated 18 models and found performance degrades as input length grows — even on simple tasks — and that topically related distractors degrade it further. Long contexts are not free memory; they are noise accumulation.
Control: context engineering. Retrieve only what the step needs, compress or summarize earlier context, and restate the current task close to the call that needs it.
Signal: task success rate vs. input-token length. If success drops as context grows, rot is active.
Failure mode 4: unverifiable output
The agent finishes, produces plausible text, and nobody can tell whether the work is actually correct — especially for code, data transforms, or multi-step operations.
Control: every task ends in an evidence artifact — a test run, a diff, a schema-validated record, a checklist the output is checked against — not a prose claim that it worked.
Signal: percentage of tasks with a passing evidence artifact. "The agent said so" is not evidence.
Failure mode 5: no human checkpoint on irreversible steps
Sending external messages, deleting records, charging money, writing to a main branch — if the agent can do these ungated, one bad reasoning chain becomes an incident.
Control: approval gates before irreversible or high-stakes actions, plus least-privilege credentials so a confused agent cannot reach what it should not touch.
Signal: count of irreversible actions executed without approval. That number should be zero, always.
When not to use an agent at all
The same Anthropic guidance makes the counterpoint: for many applications, a single model call with good retrieval and in-context examples is enough. Agents earn their complexity on open-ended problems where the steps cannot be predicted in advance. If a deterministic script meets the success criterion, it is the more reliable and cheaper choice.
The pattern
Reliable agent systems are not the ones with the best model. They are the ones where every failure mode is named, every named failure has a control, and every control has a signal someone reads. That is a reliability engineering discipline applied to a probabilistic component — the same discipline that made distributed systems operable.
Sources: Anthropic, "Building effective agents" (anthropic.com/engineering/building-effective-agents); Google Cloud, "Choose a design pattern for your agentic AI system" (docs.cloud.google.com); Chroma, "Context Rot" report, July 2025 (research.trychroma.com/context-rot). This is an engineering guide, not affiliated benchmarks.
A longer version of this guide, with the full failure-mode table and a reference implementation on the open-source Kortix/Suna agent stack, lives on the Kortix Blog.
Top comments (0)