Skip to content

drewmattie-code/Agent-Failure-Attribution

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 

Repository files navigation

Agent Failure Attribution

When the agent breaks, which layer dropped the ball?

License: CC BY 4.0 Status: v1.0 Part of: The Spine

"The AI broke" is not an attribution. It is the absence of one — and it is how most agent postmortems still end. The model gets blamed, a prompt gets tweaked, and the same class of incident comes back a month later wearing a different stack trace.

This repository is the operator-facing failure-attribution dictionary for AI agent estates: a ten-way taxonomy that assigns every agent failure to the architecture layer that owns it, a postmortem template built around that attribution, and crosswalks that map the two major academic failure taxonomies — MAST and Microsoft Research's AgentRx — onto layers a production team can actually staff, budget, and fix.

The layers are the nine specifications of the Spine, the vendor-neutral architecture catalog for running AI agents in production. Five are public: PDS, ACS, ESF, AGS, DCS.


The ten-way dictionary

When an agent system misbehaves, the failure is one of ten kinds, and each kind has exactly one owning layer:

# The failure Owned by Layer "It looked like…"
1 Bad tool data PDS Progressive Discovery Spine Wrong supplier ID, stale internal cache, missing record, malformed tool call
2 Bad world data ESF External Signal Fabric Expired signal, mis-tagged advisory, broken feed adapter
3 Bad reasoning ACS (planner) Adversarial Coordination Spine Plan unsupported by the available signals
4 Bad evaluation ACS (evaluator) Adversarial Coordination Spine Rubber-stamped contract violation, "verified" work that wasn't
5 Bad scoring CRI Composite Risk Index One mystery number, no confidence band, no method declared
6 Bad governance AGS Agent Governance Spine An action executed that policy should have denied — or a deny nobody can explain
7 Bad continuity DCS Durable Context Spine Lost state, repeated work, a false "task complete" that survived into the next session
8 Bad grounding GDS Grounded Data Spine Two agents give two answers to one metric; an agent saw data its user may not see
9 Bad or missing registry ARS Agent Registry Spine A shadow agent nobody registered; a tool nobody knew was reachable
10 Bad or unbounded execution SRS Sovereign Runtime Spine Runaway loop, unattributed action, an agent outliving its task

The rule that makes the dictionary work: every failure gets exactly one primary owner. Secondary layers may share remediation, but a postmortem that ends with two owners has not finished attributing.


Crosswalk 1 — MAST → the Spine

MAST (Cemri et al., "Why Do Multi-Agent LLM Systems Fail?") is the most-cited academic taxonomy of multi-agent failures: 14 failure modes in 3 categories, derived from 1,600+ annotated execution traces. It tells you what went wrong in the trace. This crosswalk adds who owns the fix.

Frequencies below are MAST's reported share of failures in its corpus.

MAST mode Freq. Primary owner Secondary Why
FM-1.1 Disobey task specification 11.8% ACS AGS The negotiated contract between agents was violated and the evaluator let it through. If the disobeyed rule was a governance rule, the AGS gate should have made the action structurally impossible.
FM-1.2 Disobey role specification 1.5% ACS Role separation (planner / generator / evaluator) is a structural property of the coordination layer, not a prompt aspiration.
FM-1.3 Step repetition 15.7% ACS DCS In-session loops are a coordination failure. Repetition across sessions means durable state failed to record what was already done.
FM-1.4 Loss of conversation history 2.8% DCS The textbook continuity failure: state that should have survived, didn't.
FM-1.5 Unaware of termination conditions 12.4% ACS DCS Termination criteria belong in the coordination contract; the verification-gated completion ledger is the backstop.
FM-2.1 Conversation reset 2.2% DCS An unexpected reset is a durable-state failure regardless of which agent triggered it.
FM-2.2 Fail to ask for clarification 6.8% ACS Contract negotiation must surface missing information before execution; proceeding on guesswork is a coordination failure.
FM-2.3 Task derailment 7.4% ACS Plan drift the evaluator should have caught.
FM-2.4 Information withholding 0.85% ACS An inter-agent contract violation: the information existed and was not communicated.
FM-2.5 Ignored other agent's input 1.9% ACS Same surface: coordination that discards signal.
FM-2.6 Reasoning–action mismatch 13.2% ACS AGS The action didn't follow from the stated reasoning. When the mismatched action touches the world, the deterministic gate is the last line.
FM-3.1 Premature termination 6.2% ACS DCS The evaluator accepted "done" too early; a verification-gated ledger refuses unverified completion by construction.
FM-3.2 No or incomplete verification 8.2% ACS The checker was missing or structurally able to skip the check — the exact failure ACS's separated evaluator exists to prevent.
FM-3.3 Incorrect verification 9.1% ACS The checker agreed with the maker. A checker that cannot disagree is not a checker.

What the crosswalk shows: MAST's corpus maps almost entirely onto two of the nine layers — coordination (ACS) and continuity (DCS). That is not a weakness of MAST; it studied multi-agent research systems, and coordination is exactly their failure surface. It is a measurement of how much of a production estate the research benchmarks never touch.


Crosswalk 2 — AgentRx → the Spine

AgentRx (Microsoft Research) classifies failed agent trajectories into 9 mutually exclusive failure categories, derived by grounded-theory analysis of 115 manually annotated failures across τ-bench, Flash, and Magentic-One.

AgentRx category Definition (verbatim) Primary owner Secondary Why
Plan Adherence Failure "Ignored required steps / did extra unplanned actions" ACS AGS Contract violation first; if the extra action reached the world unchecked, the gate failed too.
Invention of New Information "Altered facts not grounded in trace/tool output" GDS ESF Grounding failure: the answer must be traceable to canonical, entitled data — or to a provenance-stamped external signal.
Invalid Invocation "Tool call malformed / missing args / schema-invalid" PDS The tool surface owns its contract. Schema-invalid calls mean the discovery layer handed the agent a tool without an enforceable interface.
Misinterpretation of Tool Output "Read tool output incorrectly; acted on wrong assumptions" PDS ACS Ambiguous output is a tool-contract failure before it is a reasoning failure. Typed, disambiguated outputs shrink this class structurally.
Intent–Plan Misalignment "Misread user goal/constraints and planned wrongly" ACS Planning discipline: the negotiated contract exists to pin the goal before execution.
Under-specified User Intent "Could not proceed because required info wasn't available" ACS The coordination layer must surface the gap and ask — MAST's FM-2.2 is the same failure seen from the other side.
Intent Not Supported "No available tool can do what's being asked" PDS ARS Discovery should answer "there is no such tool" honestly — which requires a registry that actually knows what exists.
Guardrails Triggered "Execution blocked by safety/access restrictions" AGS The gate fired. Attribution decides whether the deny was correct (working as designed) or the policy was wrong (governance debt) — either way, AGS owns the surface.
System Failure "Connectivity/tool endpoint failures" SRS ESF The execution substrate owns infrastructure failure — degraded modes, retries, bounded blast radius. A failing external feed is ESF's adapter surface.

What the benchmarks can't see

Stack the two crosswalks and count: the academic corpora concentrate on ACS, DCS, PDS, and touch GDS, AGS, SRS once each. Three layers receive no coverage at all:

  • CRI (bad scoring) — research benchmarks don't ship mystery risk scores to executives.
  • ARS (bad or missing registry) — a 115-trajectory benchmark has no shadow agents; a Fortune 500 running thousands of agents does.
  • ESF (bad world data) — τ-bench doesn't make decisions on expired port-congestion signals. Supply chains do.

That is the honest reading of the literature: the taxonomies are excellent for what the benchmarks measure, and the benchmarks structurally cannot measure the failure surfaces that only exist at production scale — entitlements, registries, external signals, risk scores, runtimes. Those five surfaces are where "the AI broke" does the most financial damage, because nobody is even instrumented to name them.


The postmortem template

POSTMORTEM-TEMPLATE.md is the operational half of this repository: a fill-in incident report whose core is the attribution walk — nine questions, asked in order, that end at exactly one owning layer. It includes the symptom classification (dictionary + MAST/AgentRx codes), the SLA impact line (which target from the owning spec's table was violated), and a prevention section that asks the only question that stops recurrence: what makes this failure structurally impossible now?

How to adopt

  1. Next incident, run the walk. Use the template as-is; it assumes nothing about your stack.
  2. Name the layer in the incident channel, not the model. "That's a DCS failure" changes who gets paged and what gets budgeted.
  3. Track attribution over quarters. The distribution tells you which spec to implement next — your estate's real gap, measured on your own incidents.

Part of the Spine

This dictionary is the meta-artifact of the Spine — nine versioned specifications for running AI agents in production, each owning one failure surface. Public specs: Progressive Discovery Spine · Adversarial Coordination Spine · External Signal Fabric · Agent Governance Spine · Durable Context Spine. CRI, GDS, ARS, and SRS are private; normative summaries available on request.

Sources

  • Cemri et al., Why Do Multi-Agent LLM Systems Fail? — the MAST taxonomy: 3 failure categories, 14 failure modes, 1,600+ annotated traces. arXiv:2503.13657. Frequencies quoted from the paper.
  • Microsoft Research, Systematic debugging for AI agents: introducing the AgentRx framework — 9 failure categories over the 115-trajectory AgentRx Benchmark (τ-bench, Flash, Magentic-One). Blog post. Category definitions quoted verbatim.
  • The Spine specifications and per-layer SLA tables: saasquach.ai/spine.

Versioning

  • v1.0 (2026-07-16): initial public release — ten-way dictionary, MAST and AgentRx crosswalks, coverage analysis, postmortem template.

License

CC BY 4.0. Cite it, fork it, put it in your runbook — attribution appreciated. Authored by Drew Mattie, SaaSquach AI Labs (a division of Charles & Roe Inc.).

About

When the agent breaks, which layer dropped the ball? Operator-facing failure-attribution taxonomy for AI agent estates: ten-way dictionary, MAST + AgentRx crosswalks, postmortem template. Part of the Spine catalog.

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors