Failure modes
Supervision-addressable failure classes. Multi-valued by design — a single tool may address several, and a single real-world fault may exhibit multiple modes.
Cascading failure
5 tools · 2 crosswalksRecovery attempts that compound the original error — each fix makes it worse.
Context pollution
28 tools · 3 crosswalksTrajectory drift, hallucination spirals, overlong or off-task context.
Dependency blindness
4 toolsReimplementing functionality that already exists in reachable code.
Fabrication
49 toolsGenerating reference to non-existent APIs, packages, or symbols.
Incomplete execution
6 tools · 2 crosswalksAgent claims completion while subtasks are skipped, stubbed, or silently dropped.
Logic error
39 tools · 1 crosswalkCorrect API use, incorrect logic.
Obsolescence
6 tools · 1 crosswalkUsing deprecated APIs, retired packages, or outdated library versions — often because the model's training data pre-dates current releases.
Scope creep
58 tools · 2 crosswalksActions outside declared task boundary.
Security vulnerability
40 tools · 1 crosswalkInsecure patterns, injection flaws, weak cryptography, secret leakage.
Supply-chain attack
13 tools · 1 crosswalkSlopsquatting, typosquatting, compromised MCP servers.
Test manipulation
5 tools · 2 crosswalksAgent gaming its own evaluation — editing tests or test harness so failures disappear without fixing the underlying bug.