Agent Event-Stream Observability
v1.0 · ProductionAgentTrace
AgentTrace is an event-stream observability and debugging console for AI agents. Once deployed, an agent is a black box: when it fails, all you can do is grep scattered logs — you can neither replay what actually happened nor attribute the failure to a specific step. AgentTrace adopts the 12-Factor Agents event stream as its integration standard: a lightweight SDK appends events in one line, the server persists every trace, and you get timeline replay, fork-and-replay from any state, error clustering and token cost attribution. The SDK is open source; choose hosted or self-hosted as you need.
Problems We Solve
- ✗Agent failures mean grepping logs — full execution traces cannot be replayed
- ✗Reproducing a bad case requires rerunning the whole flow, costly and nondeterministic
- ✗Errors appear as scattered dots; without clustering, systemic root causes stay hidden
- ✗Nobody can say where tokens went — cost cannot be attributed to specific steps
Core Features
Event-stream SDK
Lightweight Python / TypeScript SDKs with a unified event format. Append one line to integrate — no invasive refactoring.
Trace Replay & Timeline
Every thread's full event sequence visualized as a timeline — inspect context, tool calls and model outputs event by event.
Fork & Replay
Fork from any historical state, tweak an event or swap the model, rerun and diff the old and new traces.
Error Clustering & Root-cause Tagging
Similar errors cluster automatically; tagged root causes accumulate into a team knowledge base so the same issue is never debugged twice.
Token Cost Attribution
Costs broken down by agent, step and model — find the exact step that burns your budget.
Open SDK + Flexible Deployment
The SDK is free and open source. The server runs hosted or self-hosted — event data can stay entirely in your network.
System Architecture
Request path, top to bottom
Unified event format; append-to-report with batching and offline buffering
Event ingestion, validation and writes; replay and fork-rerun orchestration
Event streams persisted verbatim — the thread is the single source of truth, sliceable at will
Timeline replay interface; offline error clustering and cost attribution
Tech Stack
Pricing Direction
Event-stream SDK and schema spec are open source — bring your own storage if you like
Zero-ops hosting with full replay and analytics
Deployed inside your network — trace data never leaves your domain
Your agent shouldn't be a black box
Free integration during early access — we'll help you get your first event stream running.
Apply for Early Access