Observability
Attach loggers, tracers, and cost guards to any runtime — no code changes beyond adding an observer.
Agents fail in ways that are hard to reproduce: the model chose a wrong tool, a retrieval returned stale data, a run silently blew past its budget. Observability gives you a structured record of every LLM call, tool execution, and agent step so you can answer "what happened and why" without guessing.
#Loggers + tracers
consoleLogger— local dev.langsmith— LangSmith observer.opentelemetry— OTel observer.langfuse— Langfuse observer (@agentskit/observability-langfuse).createTraceTracker— low-level span lifecycle.
#Local trace viewer
createFileTraceSink+renderTraceViewerHtml— offline Jaeger-style HTML. Recipe.
#Devtools server
createDevtoolsServer+toSseFrame— live feed for any devtools UI. Recipe.
#Cost + tokens
costGuard— hard $ ceiling per run.approximateCounter,createProviderCounter— token accounting.
#Audit
createSignedAuditLog— hash-chain + HMAC tamper-evident log. Recipe.
#Related
- Package: @agentskit/observability
- Package: @agentskit/observability-langfuse
- For agents: observability
- For agents: observability-langfuse
- Evals · Security
Explore nearby
- PeerLoggers + tracers
Attach console, LangSmith, or OpenTelemetry observers to any runtime — mix and match, all receive the same event stream.
- PeerTrace viewer
Generate a self-contained HTML trace file from any run — inspect spans offline without a tracing backend.
- PeerDevtools server
Expose a live SSE stream of agent events so any browser-based dashboard can display them in real time.
On-call runbooks
First-response playbooks for the four most common AgentsKit production incidents — LLM provider outage, tool flapping, cost spike, prompt injection.
Loggers + tracers
Attach console, LangSmith, or OpenTelemetry observers to any runtime — mix and match, all receive the same event stream.