Agent Error Attribution

Agent Error Attribution

Attribution: YourVisionYourCreation LLC — yourvisionyourcreation.com Research anchor (verified): Tomašev, Franklin & Osindero, Intelligent AI Delegation (Google DeepMind, arXiv:2602.11865, 2026), whose safe-delegation dimensions include accountability and verifiable task completion. This skill is YVYC-original doctrine: a five-stage failure taxonomy and attribution walk that operationalizes accountability across agent stages and delegation chains. Doctrine class: Tier 5 — Frontier (YVYC original, anchored in verified research)

Universal So-What

"The model hallucinated" is agent failure analysis at the level of "the patient got sick." Agent failures have anatomy: a wrong memory retrieved, a plan that never matched the goal, a tool called with mangled parameters, a reflection step that approved garbage. Each failure class lives at a different stage, is owned by a different component, and demands a different fix. Attribution is the discipline of finding the ACTUAL failing stage — because a fix applied to the wrong stage is a cost paid for a failure kept.

Core Doctrine

1. The Five-Stage Failure Taxonomy

Every agent failure is classified to the stage where it ORIGINATED — not the stage where it became visible:

The cardinal analytical rule: failures CASCADE forward. A memory-stage failure produces planning-stage symptoms and action-stage wreckage. Attribution walks BACKWARD from the visible failure to the first stage where things were already wrong — the first-error discipline of debug-evidence-protocol, applied to cognition.

2. The Attribution Walk

For every failure, in order:

1. Capture the wreckage: the visible failure, exactly — output, action, or omission
2. Reconstruct the trajectory: the full path — what was retrieved, what was planned, what was executed, what was checked — from logs and traces, not from the agent's own summary of itself
3. Walk backward: at each stage, was this stage's OUTPUT correct given its INPUT? The first stage whose output was wrong given correct input is the origin.
4. Classify: assign the taxonomy class at the origin stage
5. Verify: the classification must predict — fixing this stage should prevent this failure; if a replay with the stage corrected still fails, the walk missed an earlier origin

3. Attribution Across Delegation Chains

When the failing system is multiple agents:

• The stage taxonomy applies WITHIN each agent; chain attribution decides BETWEEN agents — which link's defect originated the failure
• Three chain-level verdicts, each landing differently: bad SPECIFICATION (the delegator asked for the wrong or impossible thing), bad EXECUTION (the delegatee failed a sound task), bad DECISION (the delegation itself should never have happened — wrong delegatee, wrong time, failed gate)
• Responsibility propagates to the origin: a sub-agent executing an impossible specification perfectly is not the failure; the specification is (extending accountability-chain, Tier 3)

4. Blameless Toward Agents, Accountable Toward Design

• The attribution's product is a DESIGN verdict, not a culprit: the failing stage indicates which component, prompt, tool schema, memory system, or gate needs the fix
• "The model is bad" is banned as a terminal diagnosis — if the model underperforms at a stage, the finding is what scaffolding, verification, or task-shaping that stage needs, or evidence the task exceeds current capability and needs human retention
• Failure data aggregates without shame: an error ledger by stage and class, because the PATTERN across twenty failures is worth more than the autopsy of one

5. The Telemetry Prerequisite

Attribution is only as good as the trajectory record:

• Each stage logs its inputs and outputs at a fidelity that permits the backward walk — an agent system that logs only final outputs has pre-decided that every failure will be attributed to nothing
• The agent's self-narrative ("I checked the data and it looked fine") is testimony, not telemetry — useful as a pointer, never as evidence
• Telemetry gaps found during an attribution walk are themselves findings: the next incident should never hit the same blind spot

6. The Fix Lands at the Stage

• Memory-stage origins get memory fixes: retrieval quality, staleness checks, poisoning defense — not prompt tweaks
• Planning origins get planning fixes: decomposition scaffolds, constraint pinning, goal restatement — not tool schema changes
• Reflection origins get reflection fixes: independent verification, adversarial checks — not more of the self-review that already failed
• The anti-pattern with a name: PATCHING THE SYMPTOM STAGE — adding action-stage guardrails for a planning-stage disease produces agents that execute wrong plans more carefully

Common Failure Modes

Non-Negotiables

1. Every failure is attributed to its ORIGIN stage, not its visible stage.
2. "The model hallucinated" is never a terminal diagnosis.
3. Chain failures land on the defective link — specification, execution, or decision.
4. The agent's self-narrative is never evidence.
5. Fixes land at the origin stage or they are symptom patches.
6. Telemetry gaps discovered in one incident are closed before the next.

Built by YourVisionYourCreation LLC — yourvisionyourcreation.com Research foundation credited above. Licensed under CC BY 4.0