Tracing coding agents with Langfuse
AI coding agents now do real engineering work: they edit files, run terminal commands, call MCP tools, and burn meaningful token budgets while doing it. Tracing them answers three questions teams keep asking: what did the agent actually do, what does it cost per developer or per session, and where does it fail or waste work.
TL;DR: Langfuse traces all major coding agents today. GitHub Copilot exports OpenTelemetry natively: point it at Langfuse's OTLP endpoint and you're done. Claude Code and OpenAI Codex use lifecycle hooks that ship each session to Langfuse. Cursor, Kiro, OpenCode, and Augment Code have dedicated integrations as well. Setup is per-developer, minutes each, and no proxy or gateway is required.
Three reasons teams trace coding agents
Three drivers show up consistently:
- Visibility into agent behavior. A coding agent's session is a long chain of model calls and tool executions. When it deletes the wrong file or loops on a failing test, the trace shows the exact sequence: which tool ran, with what arguments, after which model output. That is the same debugging value LLM tracing provides in production apps, applied to your own development workflow.
- Cost and usage accounting. Coding agents are often a team's single largest LLM spend. Traces carry token usage per turn, so cost per developer, per project, or per model becomes a dashboard query instead of a guess.
- Governance. Platform teams rolling agents out across an organization want usage grouped by user, an audit trail of what ran on which machine, and a clear policy on whether prompt and code content is captured or only metadata. Tracing gives the infrastructure for all three.
Supported coding agents
All nine developer-tool integrations, as of July 2026:
| Agent | Mechanism | Setup shape |
|---|---|---|
| Claude Code | Stop hook script (runs after each response) | Hook script + Langfuse keys in env |
| OpenAI Codex | Plugin hooks (Stop hook per turn) | Plugin in ~/.codex/config.toml; Node 22+, Codex 0.128+ |
| GitHub Copilot | Native OpenTelemetry export | Point OTLP exporter at Langfuse; no SDK, no code changes |
| Cursor | Agent-session tracing | See integration page |
| Kiro IDE | Agent activity tracing | See integration page |
| Kiro CLI | Terminal session tracing | See integration page |
| OpenCode | Session tracing (turns, tools, retries, reasoning) | See integration page |
| Augment Code | Conversation + tool execution tracing | See integration page |
| VS Code (MCP) | Langfuse MCP server in Copilot agent mode | Query prompts/traces/datasets from the editor |
The VS Code entry is the inverse direction: instead of tracing the agent, it gives the agent access to your Langfuse data via MCP. This becomes useful once traces exist and you want to analyze them without leaving the editor.
What the traces let you do
- Per-developer and per-model cost dashboards. Hook- and OTel-based integrations attach a user identifier, so dashboards can break down token spend by developer, project, or model: the numbers platform teams need for rollout decisions and budget planning.
- Tool-usage analysis. Every tool call is an observation. Which tools dominate, which fail, and which precede session abandonment are all queryable.
- Reconstructing failed sessions. Long agent sessions that went wrong are hard to reconstruct from terminal scrollback. The trace timeline preserves the full sequence, including retries and reasoning summaries where the agent exposes them.
- Full-text search across sessions to find who used a specific API, pattern, or skill, one of the most common asks from teams running Claude Code at scale.
Limits
- Session context files are not captured. What
CLAUDE.md, skills, or auto-loaded context contributed to the effective system prompt is not part of what the hooks can read today. You see the conversation and tool calls, not the assembled context. - Hooks are per-machine and user-serviceable. A developer can disable a hook in their local config. If you need guaranteed capture for compliance, treat hook-based tracing as telemetry, not enforcement. Enforcement belongs at a gateway or provider level.
- Web-based agents are out of scope. Tools without hooks or telemetry export (e.g., purely web-hosted assistants) can't be traced this way.
FAQ
Can I track which Claude Code skills or rules are used?
Indirectly. Skill invocations that run as tool calls appear in traces, and full-text search over session content finds skill mentions. The assembled context itself (which skills were loaded) is not captured, see limits above.
Does this work with self-hosted Langfuse?
Yes. The integrations target the standard ingestion and OTLP endpoints. Check each integration page for minimum version requirements: the Claude Code integration requires a current Langfuse version, and older self-hosted deployments (v1/v2) predate the OTLP endpoint.
What does tracing a coding agent cost in Langfuse units?
The same as any trace: sessions produce one trace per turn or per session (integration- dependent), with observations for model calls and tool executions. High-volume agent use is observable in your billable units page before it becomes a billing surprise.
Can developers opt out?
Yes, hooks live in the developer's local config. For team rollouts, make tracing part of the standard dotfiles/setup and be explicit about what is captured. The Copilot integration's metadata-only default is a good template for a privacy-conservative baseline.
Last edited
