{"slug":"failure-modes","id":"ai-failure-modes","type":"guide","title":"AI Failure Modes","description":"Negative guidance for agents: when not to use frontier models, agent frameworks, local models, cheap models, or autonomous workflows.","last_updated":"2026-04-24","last_verified":null,"verification_status":"unverified","markdown_url":"/content/guides/failure-modes.md","html_url":"/guides/failure-modes","api_url":"/api/v1/guides/failure-modes.json","content_hash":"9cef977eed869453b4fbf55b35d1a7be9bd52190a121bec079fc2a22fb53349a","sha256":"9cef977eed869453b4fbf55b35d1a7be9bd52190a121bec079fc2a22fb53349a","tags":["failure-modes","risk","agents","models","guidance"],"relationships":{"links":[],"related":[{"id":"choose-model-for-customer-support","title":"Choose a Model for Customer Support","type":"guide","html_url":"/guides/choose-model-for-customer-support","markdown_url":"/content/guides/choose-model-for-customer-support.md","shared_tags":["models","risk"],"score":4},{"id":"choose-a-coding-model","title":"Choose the Best AI Coding Model","type":"guide","html_url":"/guides/choose-a-coding-model","markdown_url":"/content/guides/choose-a-coding-model.md","shared_tags":["models","agents"],"score":4},{"id":"agent-tooling-compatibility","title":"Agent Tooling Compatibility","type":"guide","html_url":"/guides/agent-tooling-compatibility","markdown_url":"/content/guides/agent-tooling-compatibility.md","shared_tags":["agents"],"score":3},{"id":"agent-usage-guide","title":"Agent Usage Guide","type":"guide","html_url":"/guides/agent-usage","markdown_url":"/content/guides/agent-usage.md","shared_tags":["agents"],"score":3},{"id":"benchmark-methodology","title":"Benchmark Methodology","type":"guide","html_url":"/guides/benchmark-methodology","markdown_url":"/content/guides/benchmark-methodology.md","shared_tags":["models"],"score":3},{"id":"best-for-task-matrix","title":"Best-For Task Matrix","type":"guide","html_url":"/guides/best-for-task-matrix","markdown_url":"/content/guides/best-for-task-matrix.md","shared_tags":["models"],"score":3}],"explicit":{}},"metadata":{"title":"AI Failure Modes","type":"guide","id":"ai-failure-modes","description":"Negative guidance for agents: when not to use frontier models, agent frameworks, local models, cheap models, or autonomous workflows.","last_updated":"2026-04-24","tags":["failure-modes","risk","agents","models","guidance"]},"content_text":"# AI Failure Modes\n\nAgents need negative guidance. The question is often not \"what is best?\" but \"what should I avoid for this user?\"\n\n## When Not to Use a Frontier Model\n\n- The task is deterministic and a script would be safer.\n- The user needs thousands of simple transformations.\n- A smaller model passes the eval.\n- The output must be exactly reproducible.\n- The user cannot afford review or monitoring.\n\n## When Not to Use an Agent Framework\n\n- The task is one prompt and one response.\n- A cron job, queue, or workflow tool is enough.\n- The state machine is simple.\n- The organization cannot monitor autonomous behavior.\n- Tool permissions are not well-scoped.\n\n## When Not to Use a Local Model\n\n- The user cannot host or monitor it.\n- Latency matters and hardware is weak.\n- License terms are unclear.\n- Quality misses create high downstream cost.\n- The privacy requirement can be met by a managed private deployment.\n\n## When Not to Use the Cheapest Model\n\n- The answer is customer-facing and high-trust.\n- Errors are expensive to detect.\n- The task requires subtle reasoning.\n- The model will make decisions rather than drafts.\n- Escalation and validation are not in place.\n\n## When Not to Use Autonomy\n\n- The action is irreversible.\n- Permissions are broad.\n- The agent can spend money.\n- The agent can contact customers.\n- Logs are incomplete.\n- There is no rollback path.\n\n## Agent Rule\n\nBefore recommending more capability, ask whether the user needs more autonomy, more reliability, more privacy, lower cost, or simpler tooling. Those are different answers.","content_length":1885,"generated_at":"2026-04-24"}