1 in 20 AI agent requests fail in production right now — silently.
The system keeps running. The output looks correct. Nobody notices until a customer complains, a database is corrupted, or an audit finds something three weeks later.
76% of AI agent deployments fail within 90 days. Not because the models are bad. Because nobody can see what the agent is doing while it's doing it.
Gartner says 40% of enterprise apps will have AI agents by end of 2026. The same Gartner research says 40% of those projects will be cancelled by 2027 — specifically because of monitoring gaps.
The problem is not the model. The problem is that an agent that confidently fails is indistinguishable from an agent that correctly succeeds — unless you have a layer watching the reasoning, not just the output.
That layer didn't exist.
Until now.
AgentWatch is the first production observability layer built specifically for AI agent reasoning — not just outputs.
It sits between your agent and the world. It watches every action, scores every reasoning step with an independent model, blocks dangerous commands before they run, and gives you a full replay of exactly what happened and why.
57% of organizations run AI agents in production. Observability is the lowest-rated part of their stack. Current tools were built for single LLM calls — not multi-step agents that fail across 14 distinct failure modes.
AgentWatch was built for the agent era.
| 🧠 Reasoning Auditor | Independent LLM scores every reasoning step — not just the output |
| 🛡️ Safety Engine | Blocks dangerous commands before they execute |
| 📊 Live Dashboard | Real-time trace of every action your agent takes |
| ⏪ One-Click Rollback | Git-backed checkpoints at every step |
| 💾 Persistent Memory | Cross-session episodic, semantic, and procedural memory |
| 💰 Cost Tracker | Per-session token budget with live spend alerts |
| 🔔 Alerting | Slack + PagerDuty when confidence drops or actions are blocked |
| 📋 Compliance | GDPR/HIPAA audit exports, RBAC governance |
| 🔌 Universal | Claude Code, LangChain, AutoGPT, OpenClaw — no rewrites |
pip install agentwatch
docker compose up -dDashboard → http://localhost:3000 API → http://localhost:8000/docs
AgentWatch wraps your existing agent. You change nothing.
agentwatch watch "Build me a REST API"from agentwatch.adapters.langchain import AgentWatchCallbackHandler
handler = AgentWatchCallbackHandler()
agent = AgentExecutor(agent=..., callbacks=[handler])from agentwatch.adapters.autogpt import AutoGPTAdapter
adapter = AutoGPTAdapter(session_id="session-1")
await adapter.on_action(action)from agentwatch.adapters.openclaw import OpenClawAdapter
adapter = OpenClawAdapter(session_id="session-1")
await adapter.on_skill_execution(skill_name, payload)This is what nobody else has built.
Every agent scores its own work. And it almost always thinks it did well — even when it didn't. The confidence is the problem, not the failure.
AgentWatch deploys an independent model — architecturally separate, no access to the agent's reasoning trace — whose only job is to find failure before the next action fires.
from agentwatch.reasoning.auditor import ReasoningAuditor
auditor = ReasoningAuditor()
result = await auditor.score_step(step)
print(result.confidence) # 0.0 – 1.0
print(result.hallucination_risk) # low / medium / high
print(result.goal_drift) # True if agent is off-taskWhen confidence drops below your threshold, AgentWatch holds the next action and fires an alert. Not after the damage. Before it.
from agentwatch.core.safety import SafetyEngine
engine = SafetyEngine()
result = await engine.check_event(event)
if result.is_blocked:
print(f"Blocked: {result.safety.reasons}")
print(f"Risk level: {result.safety.risk_level.value}")Blocked by default: rm -rf /, curl | bash, disk formatting,
credential exfiltration, and 40+ other critical patterns.
agentwatch rollback <session-id> --to-step 12Or click rollback in the dashboard. Every checkpoint is a full filesystem snapshot backed by git. Irreversible actions become reversible.
GET /api/v1/sessions
GET /api/v1/sessions/{id}/replay
GET /api/v1/sessions/{id}/confidence
GET /api/v1/sessions/{id}/checkpoints
POST /api/v1/sessions/{id}/rollback
GET /api/v1/safety/blocked
GET /api/v1/dashboard/summary
WS /ws/events
- Backend — FastAPI, PostgreSQL, Redis, Celery
- Frontend — Next.js, Tailwind, Recharts, WebSockets
- Infra — Docker Compose, GitHub Actions CI
- Telemetry — OpenTelemetry compatible
✅ 47/47 tests passing
✅ docker compose up — zero errors
✅ API live at localhost:8000
✅ Dashboard live at localhost:3000
✅ Claude Code, LangChain, AutoGPT, OpenClaw adapters working
Apache 2.0