๐ ๐ - Multi-agent observability for cascade failure detection, fleet-wide heartbeat monitoring, and forensicโฆ
FAQ (answers come from our records)
How do I install it?
We did not find an install command for this repository (our install coverage is only about 4%). Check its README or releases page.
Who is it for?
We filed it under these tags: MCP ๆๅก. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
first catalogued 2026-10-14๏ผlast seen 2026-10-16ใThere is no official certification or quality rating here โ we record sources and dates only.
agentwatch
ไปฅไธๆๅฝ่ช่ฏฅไปๅบ็ README๏ผ็ๆฎต๏ผ ยท ๅจ GitHub ่ฏปๅฎๆด README โ
Your agent swarm crashed at 2am. You have logs from 10 agents and no idea which one started the cascade. AgentWatch tells you.
It tracks heartbeats, links actions across agents, walks backward from any failure to the root cause, and replays the full sequence. Works with any agent framework (CrewAI, AutoGen, LangGraph, PocketFlow, custom). Stores everything in a local SQLite file.
Early stage. Issues and feedback welcome: https://github.com/nicofains1/agentwatch/issues
---
See it in action
No install needed:
This seeds a 5-agent fleet, triggers a cascade failure, and shows you the full trace:
---
Install
Requires Node 18+. Uses better-sqlite3 (native bindings, no external database needed).
---
Quick start
---
What it does
**Heartbeats** - Each agent calls aw.report(name, status) on a schedule. AgentWatch tracks health over time and marks agents as stale or offline based on configurable thresholds.
**Cross-agent tracing** - Actions are linked by trace ID and optional parent event ID. When agent-c fails because agent-b sent bad data that came from agent-a, the full chain is queryable.
**Cascade detection** - correlate(failureEventId) walks backward from any failure to the root cause, returning the full chain with timing and output at each step.
**Alert de-duplication** - The same alert type from the same agent within a time window collapses into one entry with an incrementing count. Severity auto-escalates: info (1x) -> warning (3x) -> critical (10x).
**Forensic replay** - replay(traceId) returns all cascade chains within a trace. Useful for post-mortem analysis when a single trace touched multiple agents.
**OpenTelemetry export** - Export traces as OTEL spans (GenAI semantic conventions). Works with Jaeger, Grafana, or any OTEL-compatible backend. Requires optional peer deps.
---
CLI
Set AGENTWATCH_DB to point to your database file. Default: agentwatch.db in the current directory.
---
MCP server
AgentWatch runs as an MCP server. Add it to your Claude Code or Cursor config:
**Claude Code** (~/.claude/claude_desktop_config.json or .claude/settings.json):
**Cursor** (.cursor/mcp.json):
This exposes 13 tools: agentwatch_dashboard, agentwatch_report_heartbeat, agentwatch_trace, agentwatch_cascade, agentwatch_replay, agentwatch_get_alerts, agentwatch_get_failures, agentwatch_get_trace, agentwatch_fleet_health, agentwatch_create_trace_id, agentwatch_alert, agentwatch_resolve_alert, agentwatch_dashboard_text.