# AgentOps

> Observability for AI agents — session replay, cost and latency tracking, and debugging for multi-step runs.

AgentOps is observability built for agents specifically: session replay of every step, tool call, and LLM call, plus cost, latency, and failure tracking. A few lines of SDK turn an opaque multi-step agent run into a timeline you can debug and a dashboard you can monitor.

Website: https://www.agentops.ai

AgentOps is an observability platform built specifically for AI agents. Agents are uniquely hard to debug — one request fans out into a tree of LLM calls, tool calls, and decisions — and AgentOps turns that opacity into a **session replay**: a step-by-step timeline of everything the agent did, with cost, latency, and errors attached.

It is aimed at developers running agents in development or production who need to see why a run went wrong, what it cost, and where it slowed down. It integrates with the major agent frameworks with minimal setup.

## Highlights

- **Session replay** — a full timeline of LLM calls, tool calls, and steps for any agent run.
- **Cost & latency tracking** — per-run and aggregate spend and timing, so regressions and runaway loops surface fast.
- **Failure analytics** — catch errors, dead-ends, and repeated tool failures across runs.
- **Framework integrations** — drop-in support for popular agent frameworks (CrewAI, AutoGen, OpenAI Agents SDK, LangGraph, and more).
- **Lightweight SDK** — a couple of lines to start capturing sessions.

## In an AI-assisted workflow

```python
import agentops
agentops.init()   # then run your agent — every step, tool call, cost, and error is captured
```

> [!TIP]
> Pair agent-specific replay (AgentOps) with general LLM observability ([Langfuse](/tools/langfuse), [Arize Phoenix](/tools/arize-phoenix)) depending on whether you're debugging the agent's control flow or the underlying model calls.

## Good to know

AgentOps offers an open-source SDK with a hosted dashboard on a freemium model (free tier plus paid plans for scale and retention). You bring your agent framework and model provider. It's most useful once an agent has enough steps that logs alone stop being readable — see [agent-reliability-reviewer](/agents/meta-orchestration/agent-reliability-reviewer) for hardening what the traces reveal.

---

_Source: https://agentscamp.com/tools/agentops — Tool on AgentsCamp._