AI agent observability.
See what your AI agents are actually doing — every decision, every tool call, every retry, every outcome. Across every framework, every model vendor, and every deployment pattern.
What AI agent observability means.
AI agent observability is the practice of capturing, storing, and querying the full execution trace of an AI agent. Not just the LLM call — the agent loop. Every decision the agent made, every tool it invoked, every retry it fired, every result it produced.
Traditional observability stops at the LLM boundary. You see the prompt and the completion; you do not see the tool that returned the wrong shape, the silent retry that ate your token budget, or the sub-agent that ended up somewhere strange. AI agent observability picks up where those tools stop.
The data lives in OpenTelemetry GenAI Semantic Conventions — the open standard for instrumenting GenAI workloads. That means one canonical trace shape across LangChain, CrewAI, AutoGen, Pydantic AI, MCP servers, and any custom framework you build next. No vendor lock-in, no re-platforming.
Why AI agent observability matters.
Five outcomes engineering teams running production agents see in the first week.
Find the failing decision in seconds
Open the trace, see the exact step that broke, see what the model saw before it answered. No log archaeology.
Attribute cost to the step that emitted it
Tokens and dollars per step, per agent, per build. Catch cost drift the day it starts.
Catch regressions before customers do
Envelope alerts on retry rate, error rate, latency, and cost-per-run — routed to oncall with the trace already attached.
Swap model vendors freely
One canonical trace shape across OpenAI, Anthropic, Bedrock, Azure OpenAI, Gemini. Vendor decisions become data, not lock-in.
Adopt without re-platforming
OpenTelemetry-native. Add Trefur as a second exporter alongside Datadog, Honeycomb, Grafana. Existing pipeline stays in place.
How Trefur does AI agent observability.
Six capabilities — built on OpenTelemetry, designed to drop into the runtime you already ship.
Agent traces
One root span per agent run. Every LLM call, tool call, retry, and outcome as a child span.
Read moreCost tracking
Tokens and dollars attributed to the step that emitted them. Roll up by agent, build, deployment, vendor.
Read moreAlerts
Envelope-based detection on retry rate, cost, latency, errors. Routed to Slack, PagerDuty, Teams, webhook.
Read moreCross-vendor coverage
One trace shape across OpenAI, Anthropic, Bedrock, Azure OpenAI, Gemini. Compare two vendors on the same workload.
Read moreSelf-hostable collector
Single Go binary inside your network. Local batching, on-disk buffering, redaction.
Read moreReplay debug
Reconstruct any failing production run with the exact prompt the model saw and the exact tool response it got back.
Read moreFive minutes to the first trace.
Drop the SDK in. Existing agent code runs unchanged. Every decision, tool call, and retry shows up in Trefur.
from trefur_observe import TrefurObserve, patch_langchain
from langchain.agents import AgentExecutor
obs = TrefurObserve.init(api_key="trf_obs_...")
patch_langchain(obs)
# Existing LangChain agent — no other changes
agent = AgentExecutor.from_agent_and_tools(...)
agent.invoke({"input": "Summarise this support ticket"})
# Every decision, tool call, and retry now appears in Trefur.Get AI agent observability today.
Free tier. No card required. First trace in under five minutes.