AgentWatch

agentwatch

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:

npx @nicofains1/agentwatch demo

This seeds a 5-agent fleet, triggers a cascade failure, and shows you the full trace:

AgentWatch Fleet Dashboard

============================================================ Agents: 5 total | 3 healthy | 1 degraded | 1 error | 0 offline

Cascade Failure (4 steps, root cause: scheduler/dispatch-batch) ============================================================ [ROOT] scheduler/dispatch-batch [ok] 15ms {"assigned_to": "fetcher"} | [ 1 ] fetcher/call-api [error] 30000ms TIMEOUT after 30000ms | [ 2 ] processor/transform [error] 120ms Error: input is null - expected array from fetcher | [FAIL] notifier/send-alert [error] 8ms Error: no processed data to report

---

Install

npm install @nicofains1/agentwatch

Requires Node 18+. Uses better-sqlite3 (native bindings, no external database needed).

---

Quick start

import { AgentWatch } from '@nicofains1/agentwatch';

const aw = new AgentWatch(); // creates agentwatch.db in the current directory

// Report heartbeats from your agents aw.report('agent-a', 'healthy'); aw.report('agent-b', 'healthy');

// Trace an action in agent-a const traceId = aw.createTraceId(); const e1 = aw.trace(traceId, 'agent-a', 'fetch-data', 'url=https://api.example.com', 'rows=150');

// Trace a dependent action in agent-b that fails const e2 = aw.trace(traceId, 'agent-b', 'process', JSON.stringify({ rows: 150 }), 'Error: out of memory', { parentEventId: e1.id, status: 'error', durationMs: 4200, });

// Walk back to the root cause const chain = aw.correlate(e2.id); console.log(chain?.root_cause); // -> { agent: 'agent-a', action: 'fetch-data', ... }

// Print fleet status console.log(aw.dashboardText());

---

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

npx @nicofains1/agentwatch demo                   # run the demo

npx @nicofains1/agentwatch dashboard # fleet health overview npx @nicofains1/agentwatch cascade # trace cascade from a failure npx @nicofains1/agentwatch failures [agent] # list recent failures npx @nicofains1/agentwatch alerts [agent] # list active alerts npx @nicofains1/agentwatch replay # replay all cascades in a trace npx @nicofains1/agentwatch mcp # start MCP server (stdio)

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):

{

"mcpServers": { "agentwatch": { "command": "npx", "args": ["@nicofains1/agentwatch", "mcp"], "env": { "AGENTWATCH_DB": "/absolute/path/to/agentwatch.db" } } } }

Cursor (.cursor/mcp.json):

{

"mcpServers": { "agentwatch": { "command": "npx", "args": ["@nicofains1/agentwatch", "mcp"], "env": { "AGENTWATCH_DB": "/absolute/path/to/agentwatch.db" } } } }

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.

---

API reference

Constructor

const aw = new AgentWatch({

db_path: 'agentwatch.db', // SQLite file path alert_window_minutes: 30, // de-dup window for alerts heartbeat_stale_minutes: 30, // when to mark agents as offline });

Heartbeats

aw.report(agent, status, context?)     // status: 'healthy' | 'degraded' | 'error' | 'offline'

aw.getLatestHeartbeat(agent) // -> Heartbeat | undefined aw.getFleetHealth() // -> AgentHealth[]

Tracing

aw.createTraceId()                                // -> string (UUID)

aw.trace(traceId, agent, action, input, output, { parentEventId?: number, status?: 'ok' | 'error', // default: 'ok' durationMs?: number, }) // -> TraceEvent aw.getTraceEvents(traceId) // -> TraceEvent[] aw.getRecentFailures(agent?, limit?) // -> TraceEvent[]

Cascade detection

aw.correlate(failureEventId)    // -> CascadeChain | null

aw.replay(traceId) // -> CascadeChain[]

Alerts

aw.alert(agent, alertType, message)

aw.resolveAlert(alertId) aw.activeAlerts(agent?) // -> Alert[]

Dashboard

aw.dashboard()      // -> DashboardOutput (structured)

aw.dashboardText() // -> string (formatted for terminal)

OpenTelemetry export

Requires optional peer deps @opentelemetry/api and @opentelemetry/sdk-trace-base.

await aw.exportTraceToOtel(traceId, { serviceName: 'my-agents' });

await aw.exportRecentToOtel(1); // last 1 hour

---

Storage

SQLite via better-sqlite3. The database file is created automatically on first use. WAL mode is on for concurrent reads.

Tables: heartbeats, trace_events, alerts.

---

License

MIT