Agent Error Classification

Cross-session analysis of errors, corrections, and failures across Claude Code & Codex history • June 2026

1,822
Total error-related messages
14
Error categories identified
1,136 / 686
Claude / Codex split
44+
Projects with errors

Error Taxonomy

Errors classified into 3 tiers by severity and actionability:

Tier 1 — Structural Errors (agent doing the wrong thing)

These indicate fundamental misalignment between what the agent did and what was wanted.

1. Misunderstood User Intent 22 occurrences (1.2%)

Agent interprets a user request literally or incorrectly, producing the wrong output entirely.

"ok when I said radio buttons, I meant a toggle"
"no i mean sounds good for you to do the short pass..."

Claude: 14 Codex: 8

2. Scope Creep / Unwanted Changes 16 occurrences (0.9%)

Agent modifies code or adds features the user never asked for, breaking existing functionality or adding complexity.

"woah wait why did you make that change to the index? everything in the old code works"
"why did you ship to a feature branch, you are explicitly not supposed to!!!"

Claude: 2 Codex: 14 — Codex 7x more prone, likely due to autonomous execution without approval gates.

3. Git Workflow Violations 22 occurrences (1.2%)

Agent ships to feature branches when told not to, restores deleted branches, or loses work during cherry-picks/merges.

"why did you ship to a feature branch, you are explicitly not supposed to!!!"
"when we did our branch merges and cherry picks, I think our skills lost a lot of the different fixes"

Claude: 16 Codex: 6

4. Hallucination / Fabricated References 7 occurrences (0.4%)

Agent references APIs, files, or capabilities that don't exist. Low frequency but high cost per incident.

Claude: 5 Codex: 2

Tier 2 — Execution Errors (agent doing the right thing badly)

Agent understood the intent but produced buggy or incomplete output.

5. Persistent / Repeated Failures 275 occurrences (15.1%)

Second largest category. User asks agent to retry something that failed previously, or reports the same issue persisting after a "fix." Indicates the agent's fix didn't address root cause.

"can you try npm run build again?"
"Ok the terminal is popping back up again, but we still have no interactivity"

Claude: 164 Codex: 111

6. UI / Visual Bugs from Agent Changes 151 occurrences (8.3%)

Agent makes UI changes that produce visual regressions — broken layouts, wrong scroll behavior, incorrect positioning, phantom UI elements.

"nope now the image looks warped, I think we just need to move the svgs up"
"the list containers extend beyond the height of the board container. There is also no scroll"

Claude: 98 Codex: 53

7. Type / API Contract Mismatches 47 occurrences (2.6%)

Frontend expects one shape, backend returns another. Includes JSON parse failures, unexpected status codes, type errors at runtime.

"Failed to load sessions: [INVALID_RESPONSE] API returned unexpected data format"
"Failed to check auth status: SyntaxError: Unexpected token '<', '<!DOCTYPE'..."

Claude: 38 Codex: 9 — Heavily Claude-skewed; Bismarck project accounts for most.

8. State Management & Persistence Issues 254 occurrences (13.9%)

Cache not invalidated, sessions not reconnecting, auth state lost on refresh, login loops. Agents struggle with stateful systems.

"when I hard refresh while keeping the session open, it does not reconnect"
"we have to log in every time even if I reconnect to a session"

Claude: 134 Codex: 120

9. Incomplete / Missing Functionality 85 occurrences (4.7%)

Agent claims work is done but user discovers missing features, skipped steps, or incomplete implementations.

"ok what about the missing features?"
"did you add all of your findings? it seems quite light"

Claude: 36 Codex: 49 — Codex slightly more prone, possibly due to less interactive verification.

10. Deployment & Infrastructure Errors 143 occurrences (7.8%)

Build failures, deploy misconfigurations, staging/prod environment issues, service connectivity problems.

"the https://www.boston4.builders/admin/subscribers page is still down with... server-side exception"

Claude: 84 Codex: 59

Tier 3 — Navigational (course corrections, not errors)

Normal direction-setting that's part of collaborative workflow. Not true agent mistakes.

11. Direction Corrections 333 occurrences (18.3%)

Largest category. Mostly "no, do X instead" — user steering rather than agent failures. High noise in this bucket.

12. Bug Fix Requests 319 occurrences (17.5%)

User asking agent to fix a bug — could be pre-existing or agent-introduced. Ambiguous attribution.

13. Error Reports 76 occurrences (4.2%)

User pasting error messages for agent to diagnose. Informational, not necessarily agent-caused.

Key Patterns & Insights

Pattern 1: Persistent Failures are the biggest real problem

At 15.1% of all error messages, "try again" / "still broken" indicates agents frequently apply surface-level fixes that don't address root causes. The fix-retry loop wastes the most user time of any category.

Pattern 2: State & UI are the hardest domains for agents

State management (13.9%) and UI visual bugs (8.3%) together account for 22% of errors. Both require runtime understanding that agents lack — they can't see the screen or observe session state. The bismarck-v0.3 project alone generated 173 error messages, mostly auth/session/terminal state issues.

Pattern 3: Codex is more prone to scope creep; Claude to type mismatches

Codex scope-creep rate is 7x Claude's (14 vs 2), likely because it runs autonomously without approval gates. Claude generates 4x more type/API mismatches (38 vs 9), concentrated in full-stack projects where it modifies both frontend and backend without cross-checking contracts.

Pattern 4: Project complexity correlates with error density

Top error-producing projects are full-stack apps with auth, real-time features, and deployment pipelines: bismarck-v0.3 (173), loadoutworks.com (111), metternich-engine (110), monorepo (90). Tooling repos like agentic-skills (59) produce fewer errors per session.

Error Taxonomy Summary

TierCategoryCount%SeverityActionability
1Misunderstood Intent221.2%HighClarification gates, confirm before executing
1Scope Creep160.9%HighStricter diff scoping, approval before extra changes
1Git Workflow Violations221.2%HighEnforce CLAUDE.md git conventions via hooks
1Hallucination70.4%HighVerify-before-recommend pattern
2Persistent Failures27515.1%Med-HighRoot-cause analysis requirement before retry
2UI/Visual Bugs1518.3%MediumScreenshot verification, browser testing
2Type/API Mismatches472.6%MediumShared types, contract-first development
2State/Persistence25413.9%MediumState diagram review before changes
2Incomplete Work854.7%MediumChecklist verification before marking done
2Deploy/Infra1437.8%MediumStaging validation, smoke tests
3Direction Corrections33318.3%LowNormal workflow — not actionable
3Bug Fix Requests31917.5%LowAmbiguous attribution
3Error Reports764.2%LowInformational

Recommendations

PriorityError ClassMitigationType
1Persistent FailuresAdd CLAUDE.md rule: "When a fix doesn't work on first try, perform root-cause analysis before attempting a second fix. Never retry the same approach."Convention
2State/PersistenceFor stateful features (auth, sessions, cache), require agent to describe the state machine before modifying code.Convention
3UI/Visual BugsEnforce browser verification for all UI changes (already in CLAUDE.md — compliance is the gap).Convention
4Scope Creep (Codex)Tighten Codex task scoping: one specific change per prompt, explicit "do not modify other files" when needed.Workflow
5Type/API MismatchesShared types between frontend/backend; agent must verify API contract before modifying either side.Convention
6Git WorkflowPre-push hook enforcing branch policy; already in CLAUDE.md but Codex doesn't always read it.Hook