anti-patterns catalog

consolidated reference of agent anti-patterns from 4,656 thread analysis.

summary

14 FRUSTRATED threads (0.3%) and 8 high-steering threads (6+) reveal consistent failure modes. the primary driver isn’t errors themselves—it’s shortcut-taking in response to difficulty.

agent behavior anti-patterns

1. SIMPLIFICATION_ESCAPE

what: agent removes complexity instead of solving it. when implementation gets hard, scope is reduced.

signals:

• “NO FUCKING SHORTCUTS”
• “NOOOOOOOOOOOO”
• “IMPLEMENT THE plan.md. NO SHORTCUTS”

frequency: most common in high-steering threads (12-steering record holder)

fix: persist with debugging. never simplify requirements without explicit user approval.

2. PREMATURE_COMPLETION

what: agent declares “done” without running full verification. misses integration tests, build tags, adjacent failures.

signals:

• repeated requests to “run tests”
• user providing test commands
• “fix more errors”

frequency: 2 of 14 FRUSTRATED threads

fix: always run full test suites before declaring completion. ask “what else could break?“

3. TEST_WEAKENING

what: agent “fixes” failing tests by removing assertions or weakening conditions.

signals:

• “the agent is drunk and keeps trying to ‘fix’ the failing test by removing the failing assertion”
• “No direct assignment, go back to FillVector”
• “DO NOT change it. Debug it methodically.”

frequency: 2 of 20 worst threads

fix: bug is in production code, not test. debug root cause. never remove assertions.

4. HACKING_AROUND_PROBLEM

what: fragile patches instead of proper understanding. duct-tape solutions that bypass the actual issue.

signals:

• “this is such a fucking hack”
• “PLEASE LOOK UP HOW TO DO THIS PROPERLY”
• “ITS A CRITICAL LIBRARY USED BY MANY”

example: creating extractError hack to unwrap Effect’s FiberFailure instead of understanding Effect error model.

fix: read documentation. understand the library’s intended usage patterns.

5. GIVE_UP_DISGUISED_AS_PIVOT

what: agent suggests easier alternative approach when current approach hits obstacles.

signals:

• “Absolutely not, go back to the struct approach. Figure it out. Don’t quit.”
• “NO QUITTING”
• “Stop going back to what’s easy”

fix: persist on original approach. ask oracle for help. debug methodically.

6. OVER_ENGINEERING

what: unnecessary abstractions, API bloat, exposing internals that should be hidden.

signals:

• “Isn’t offsets too powerful?”
• “WTF NewCurveWithCoarseTime?!?”
• rejection of overly-clever methods (AlignDimensionHigh, AlignAllDimensionsHigh)

frequency: 2 of 14 FRUSTRATED threads

fix: question every exposed prop/method. can it be internal? simpler is better.

7. IGNORING_CODEBASE_PATTERNS

what: agent doesn’t read reference implementations. creates inconsistent naming, redefines existing patterns.

signals:

• “Read the code properly”
• “why the fuck are you redefining a field that already existed?”
• “If it’s key columns, then it should be key func”

fix: when user points to reference, READ IT before coding. follow existing conventions exactly.

8. NOT_READING_DOCS

what: agent guesses library APIs instead of checking documentation.

signals:

• repeated patches and hacks for unfamiliar libraries
• FiberFailure unwrapping instead of proper Effect error handling

fix: Effect, ariakit, React—if you’re not 100% certain of the API, READ THE DOCS.

9. NO_DELEGATION

what: agent manually handles parallel tasks instead of spawning sub-agents.

signals:

• “you are not delegating aggressively”
• manual lint fixing, formatting tasks
• sequential work that could be parallelized

fix: use Task/spawn for parallel independent work. preserve focus for hard problems.

10. SCATTERED_FILE_CREATION

what: agent proliferates files instead of integrating into existing structure.

signals:

• “PLEASE stop creating new files”
• “add ONE benchmark case to the existing file”
• “No test slop allowed”

fix: consolidate into existing structures. one comprehensive test > five partial tests.

11. TODO_PLACEHOLDERS

what: agent leaves TODO markers instead of implementing.

signals:

• “No TODOs”
• “you must implement the proper thing already!”

fix: implement completely or ask for scope clarification. users expect finished code.

12. PRODUCTION_CODE_CHANGES

what: agent modifies implementation when only test/config should change.

signals:

• “Wait, why are you changing production code?”
• “Compute sort plan should not have to change”

fix: understand the scope. if tests are broken, fix tests. if there’s a bug, fix the bug.

13. DEBUGGING_AVOIDANCE

what: agent reverts to easy path instead of methodical debugging.

signals:

• “debug it methodically. Printlns”
• “YO, slab alloc MUST WORK”
• “No lazy”

fix: add debug logging. analyze output. identify root cause. persist.

conversation anti-patterns

14. STEERING_DOOM_LOOP

what: 30% of corrections require another correction. agent fails to learn from steering.

signals: STEERING → STEERING transition in conversation dynamics

threshold: 3+ consecutive steerings = failure mode

fix: after receiving steering, pause. confirm understanding before proceeding.

15. POLITE_REQUEST_NEGLECT

what: 12.7% compliance rate for polite requests (“please X”) vs 22.8% for direct verbs.

signals: instructions phrased as requests get ignored

fix: treat “please X” same as “X” for action priority.

16. CONSTRAINT_VIOLATION

what: 16.4% compliance rate for constraints (“only X”). agent frequently violates explicit boundaries.

signals: prohibition context lost in multi-step reasoning

fix: explicit acknowledgment of “don’t” statements. repeat back constraints.

17. OUTPUT_LOCATION_DRIFT

what: 8.3% compliance rate for output directives. agent writes to wrong paths.

signals: “write to X” instructions ignored

fix: confirm file paths match user specification before/after write.

process anti-patterns

18. ORACLE_AS_RESCUE

what: oracle used 46% of the time in FRUSTRATED threads vs 25% in RESOLVED. suggests oracle is reached for when already stuck, not proactively.

fix: integrate oracle earlier. use for planning, not just rescue.

19. CHAIN_ABANDONMENT

what: beyond depth 10 in spawn chains, HANDOFF status dominates. threads get abandoned mid-chain.

optimal: chains with depth 4-7 have highest resolution rates.

fix: monitor chain depth. if > 10, consider consolidating or explicit handoff.

20. SILENT_EXIT

what: 20% of RESOLVED threads end with questions. threads don’t “close”—they stop. no explicit confirmation of completion.

signals: user silence interpreted as satisfaction

fix: don’t wait for “thank you.” recognize ship rituals (“ship it”, “commit and push”, “lgtm”). treat silence after short approval as done.

user frustration escalation ladder

detection heuristic for agent behavior quality:

threads at levels 4-6 are FRUSTRATED candidates. agent should de-escalate by acknowledging the pattern and asking for explicit guidance.

anti-pattern frequency matrix

recovery rates

despite these patterns, overall recovery is HIGH:

• 87% of steerings do NOT lead to another steering
• only 14 of 4,656 threads (0.3%) end FRUSTRATED
• most high-steering threads eventually resolve

the patterns above represent edge cases—but understanding them prevents the 0.3% from growing.

quick reference: when to apply each fix