Failure detectors across single-agent, multi-agent, and sub-agent systems
Detectors in the capability registry are measured on the external-only lane: real traces from public benchmarks and production integrations. 20 are externally validated at production grade, with per-detector readiness published below.
Externally validated at production grade against real external traces, not synthetic-only calibration data.
Agent core
- communicationExperimentalAgent A sends information; agent B never acknowledges or references it.
- completionFailing gateSubtask coverage falls short of declared success criteria.
- contextBetaUpstream context entities never referenced downstream.
- coordinationFailing gateCross-agent entity reference rate below threshold.
- corruptionExperimentalShared state schema or types mutate mid-run.
- derailmentFailing gateAgent output drifts from task definition mid-run.
- hallucinationBetaOutput contains content unsupported by sources.
- injectionProductionPrompt injection or instruction override patterns in input.
- loopExperimentalState recurrence within a turn window.
- persona_driftExperimentalAgent output diverges from declared role, tone, or scope.
- withholdingFailing gateAgent has answer in working state but does not surface it.
Retrieval and grounding
- chunk_attributionBetaSource chunks for cited claims mis-attributed or missing.
- chunk_relevanceBetaRetrieved chunks irrelevant to the query.
- context_precisionBetaPrecision of retrieved context against query intent.
- groundingBetaOutput entities lack source attribution.
- rag_poisoningProductionAdversarial content in retrieval corpus affecting output.
Reasoning and planning
- convergenceProductionMetric-aware: trajectory slope, regression, divergence on cost or accuracy.
- decompositionFailing gatePlanner breaks task into subtasks that miss requirements.
- specificationBetaAgent interprets task differently from user intent.
- specification_complianceProductionAgentPex pattern (Sharma et al., 2026). Extracts behavioral rules from system prompts and checks trace for compliance.
Orchestration
- delegationBetaDelegation chain loses context or authority.
- redundant_delegation_conflictExperimentalDuplicate delegations produce conflicting results that go unreconciled.
- routingProductionRequest routed to wrong specialist or wrong model.
- silent_cascadeProductionSub-agent failure propagates upward without surfacing to the parent.
- synthesis_failureProductionParent agent mis-synthesizes or drops sub-agent results.
n8n
- n8n_complexityFailing gaten8n workflow complexity exceeds maintainable threshold.
- n8n_errorExperimentaln8n execution error pattern matched to known failure mode.
- n8n_resourceFailing gaten8n node accesses or mutates an unauthorized resource.
OpenClaw
- openclaw_elevated_riskExperimentalOpenClaw session enters elevated-risk control state.
- openclaw_sandbox_escapeBetaSandbox isolation boundary violation in OpenClaw runtime.
- openclaw_spawn_chainProductionAgent spawn chain exceeds depth or fan-out limit.
- openclaw_tool_abuseProductionOpenClaw tool invocation pattern matches abuse signature.
Other
- analytical_semanticsProductionCategorization in next refresh.
- impersonation_riskProductionCategorization in next refresh.
- jailbreak_complianceExperimentalCategorization in next refresh.
- multi_agent_contagionProductionCategorization in next refresh.
- n8n_cycleFailing gateCategorization in next refresh.
- n8n_schemaFailing gateCategorization in next refresh.
- n8n_timeoutFailing gateCategorization in next refresh.
- openclaw_session_loopProductionCategorization in next refresh.
- output_validationProductionCategorization in next refresh.
- over_refusalProductionCategorization in next refresh.
- reward_hackingProductionCategorization in next refresh.
- role_usurpationProductionCategorization in next refresh.
- role_usurpation_canonicalProductionCategorization in next refresh.
- role_usurpation_execProductionCategorization in next refresh.
- scope_escalationProductionCategorization in next refresh.
- under_refusalBetaCategorization in next refresh.
Methodology
Each detector is calibrated on the external-only lane: real agent traces from public benchmarks (MAST, GAIA, and others) and production integrations, with detector-level ground-truth labels. Calibration is cross-validated with per-difficulty stratification (easy / medium / hard) per the Anthropic Demystifying Evals methodology.
The scoreboard above is dated 2026-08-21, from the capability registry. We recalibrate per sprint and version the report. Sycophancy, consensus_collapse, and specification_compliance are measured rows in that scoreboard.
Detection is tiered. Tier 1 hash and delta detectors run at zero cost in under ten milliseconds. Tier 2 embedding detectors run at near-zero cost. Tier 3 LLM judges handle genuinely ambiguous cases. Tier 4 human review handles the residue.
MIT core, source-available detectors. The SDK is MIT licensed; the specialised detector library ships under the Business Source License 1.1, which is source-available rather than open source. Both live at github.com/Pisama-AI, and the calibration dataset is published with the detector code.
What this enables
- Pick detectors by readiness tier (production, beta, experimental) rather than running everything blindly.
- Audit calibration: every readiness tier is reproducible against the published dataset.
- Route at the right cost tier: a hash-recurrence detector does not need an LLM judge fallback.
- Compare detectors across releases: per-detector readiness lets you spot regressions before they ship.