Skip to content

Pisama

Multi-agent failure detection & self-healing platform

Pisama detects 22 failure modes in LLM agent systems, from infinite loops and hallucinations to coordination breakdowns and prompt injection. Built on the MAST taxonomy, Pisama provides production-grade observability for any multi-agent framework.

  • Get started in 5 minutes

    Install Pisama, send your first trace, and see failure detection in action.

    Quickstart

  • API Reference

    Complete REST API documentation for traces, detections, healing, analytics, and integrations.

    API docs

  • Detection Reference

    Detailed documentation for all 22 failure mode detectors with accuracy benchmarks.

    Detectors

Why Pisama?

The problem

LLM agents fail silently. A coding agent loops for 40 minutes burning tokens. A research agent hallucinates citations. A customer support agent drifts from its persona. A multi-agent pipeline drops critical context between handoffs. These failures are invisible to standard monitoring tools.

The solution

Pisama provides purpose-built failure detection for AI agent systems:

Capability Description
22 failure detectors Covering planning, execution, verification, and cross-cutting concerns
Tiered detection Cost-aware escalation from hash-based ($0.00) to LLM judge ($0.05)
Framework agnostic Works with LangGraph, CrewAI, AutoGen, n8n, Dify, OpenClaw, Claude Code
Self-healing Automated fix generation, approval workflows, and rollback capabilities
OTEL native Built on OpenTelemetry with gen_ai.* semantic conventions
Production accuracy 29 detectors at F1 >= 0.80, with continuous calibration

Supported frameworks

  • LangGraph -- State graph analysis, node-level detection
  • n8n -- Workflow validation, AI node monitoring, webhook integration
  • Dify -- App monitoring, workflow execution tracking
  • CrewAI -- Agent role validation, task delegation analysis
  • AutoGen -- Multi-agent conversation monitoring
  • OpenClaw -- Session and agent tracking
  • Claude Code -- Tool call capture and session analysis

Architecture at a glance

Trace Sources                   Pisama Platform                    Outputs
--------------                  ---------------                    -------

 LangGraph  ----+
 n8n        ----+---> Ingestion ---> Detection Engine ---> Dashboard
 Dify       ----+     Pipeline       (22 detectors)        Alerts
 CrewAI     ----+     (OTEL/         Tiered escalation     Fix suggestions
 Claude Code ---+      webhook)      LLM Judge             Self-healing
                                                            API / webhooks