AI agent failure detection

Find where an AI agent run went wrong.

Pisama analyzes AI-agent execution traces to detect failure patterns and identify the agent and step involved.

Evidence boundaryTrace evidence can locate likely failures. It does not prove every real-world outcome. Verification outside the trace depends on access to reliable system APIs.

What Pisama does

From an agent trace to a focused failure finding.

01
Ingest the trace
Accept execution evidence from an AI agent run.
02
Detect failure patterns
Run measured detectors against the trace.
03
Locate the failure
Point to the agent and step associated with the finding.

Measured detector registry

48 measured detectors, shown by readiness.

20 currently meet the production gate. The other 28 remain visible because their readiness matters.

Calibration snapshot: August 21, 2026

Production

These detectors meet the current production release gate.

analytical_semantics
convergence
impersonation_risk
injection
multi_agent_contagion
openclaw_session_loop
openclaw_spawn_chain
openclaw_tool_abuse
output_validation
over_refusal
rag_poisoning
reward_hacking
role_usurpation
role_usurpation_canonical
role_usurpation_exec
routing
scope_escalation
silent_cascade
specification_compliance
synthesis_failure

Beta

These detectors are calibrated, but not yet production-gated.

chunk_attribution
chunk_relevance
context
context_precision
delegation
grounding
hallucination
openclaw_sandbox_escape
specification
under_refusal

Experimental

These detectors are exploratory and may change.

communication
corruption
jailbreak_compliance
loop
n8n_error
openclaw_elevated_risk
persona_drift
redundant_delegation_conflict

Below release gate

These detectors currently fail the release gate.

completion
coordination
decomposition
derailment
n8n_complexity
n8n_cycle
n8n_resource
n8n_schema
n8n_timeout
withholding