Agent Failure Registry

A neutral namespace for AI agent failure modes, and a crosswalk between the taxonomies that already exist.

Not another classifier. There are at least seven of those, each with its own private vocabulary: Agent-Xray's tool_selection_bug, AgentRx's Invalid Invocation, and five more describing overlapping reality under names that share nothing. This is the mapping layer between them.

>>> import afr
>>> afr.map("agentrx", "Invalid Invocation")
[AF-0023 (exact)]
>>> afr.map("agent-xray", "spin")
[AF-0058 (exact)]
>>> afr.mode("AF-0058").layer
'harness'

The analogy is CVE, not CWE. Plenty of people have proposed taxonomies; nobody has done the boring institutional work of assigning stable ids, accepting mappings, and letting tools reference them. CVE's value was never intellectual — it was giving the whole industry a shared noun.

Why this and not a classifier

• It gets stronger as rivals ship. Every new classifier with a private taxonomy increases the need for the crosswalk. A classifier is threatened by the next classifier; a namespace is fed by it.
• Incumbents structurally cannot build it. A registry operated by one vendor is one every competitor is right to ignore. Microsoft declined the standards position for AgentRx explicitly.
• Distribution runs through other repos. You do not need an audience — you open a PR adding a mapping file to each classifier.

Install

pip install agent-failure-registry     # zero runtime dependencies; imports as `afr`

Use

import afr

# One tool's report, tagged with shared ids
labels = afr.normalize({"stuck_loop": 31, "context_overflow": 18}, "agent-xray")
print(afr.profile(labels).render())

failure profile  (source: agent-xray, n=49)
--------------------------------------------------------------------
AF-0058    63.3%  ######################              Unproductive repetition
AF-0064    36.7%  ############                        Context overflow truncation
--------------------------------------------------------------------
by layer: harness=49

That is the pitch in one output. 67% pass tells you nothing you can act on.

afr modes                              # every mode
afr show AF-0142                       # full record, plus who maps to it
afr map agentrx "System Failure"       # source category -> AF ids
afr unmap AF-0064 agent-xray           # and back, with relations inverted
afr coverage                           # how much of each taxonomy is mapped
afr gaps                               # what nobody has named yet

Two levels of identifier

Conflating them breaks the query that justifies the whole project: did Opus 5.2 regress on AF-0142 relative to 5.1? That needs modes stable across time and instances pinned to versions — so they cannot be the same identifier.

Instances are deliberately not implemented yet. Open collection only once three independent tools emit AF ids.

Relation direction

relation describes the AF mode relative to the source category: exact, broader, narrower, overlaps. afr.unmap() inverts it for you — swapping broader and narrower — because hand-authoring the reverse is how crosswalks silently corrupt.

At most one exact per source category. One wrong exact discredits the table.

Status

Pre-v0. 24 modes, all provisional; 6 source taxonomies, every one mapped from a full read of its operative definitions (classifier code, judge prompts, paper text, ground-truth annotations), with the evidence in registry/crosswalk/notes/:

No mode is an orphan: every AF id has at least one source, and the five newest (AF-0153 required step omitted, AF-0157 wrong tool selected, AF-0161 fabricated content, AF-0166 unnecessary action, AF-0170 external service failure) each have two or more. The rule that produced them: a gap becomes a mode only when a second independent taxonomy names the same thing — one mode per vendor category is how a shared namespace becomes a house style.

afr gaps lists what is still unnamed. The ones with more than one source behind them are the next candidates; the single-source ones (the whole memory family, recovery failure, sycophancy, prompt injection, external access blocks) wait for a second.

One open question the read surfaced rather than settled: Model or Harness? puts repetition, delegation and insufficient-sourcing failures on the model side because the information to avoid them was in context; AFR currently has them on harness. See notes/model-or-harness.md.

Repo layout

registry/modes/AF-####.yaml     canonical mode records (CC0)
registry/crosswalk/*.yaml       one file per third-party taxonomy (CC0)
registry/crosswalk/notes/       the full-read evidence behind each crosswalk
registry/SCHEMA.md              field semantics and relation direction
afr/                            the library (Apache-2.0), no dependencies
afr/_index.json                 compiled from YAML, committed, CI-enforced
scripts/compile.py              YAML -> index (the only place PyYAML is needed)
scripts/validate.py             integrity checks; a lying crosswalk is worse than none

Development

make check      # build + validate + test
make demo       # coverage, gaps, and a sample profile

Licensing

Registry data is CC0; the library is Apache-2.0. A namespace with an attribution clause invites legal review, and legal review is the friction that stops tools embedding it.

Governance, including the commitment to move to a neutral foundation once three independent tools map to this namespace, is in GOVERNANCE.md.

Outreach

Distribution runs through other projects' repos, not through this one. Drafted mapping-file PRs for AgentRx and Agent-Xray, and an issue for AgentDebugX, are in outreach/ — including the order to send them in, which matters more than the speed.