This week, we launched Halo, the assistant that lives inside Langfuse. Halo grew out of the infrastructure we built to support external agents using Langfuse, which we’ve been encouraging for a while.

Halo can look at what is happening in your project, work with your traces, sessions, and metrics, and take action.

What we've been using Halo for

We've been using Halo for many different things:

• Getting started with evals. Halo looks at your project, determines a useful first step, and proposes concrete signals to collect.
• Catching conversation signals. Halo can set up an evaluator that catches users who disagree, sound frustrated, ask repeated follow-up questions, or say they are happy.
• Building dashboards. Describe what you want to monitor and let Halo create the widgets for you.
• Finding ways to reduce costs. Halo can analyze your traces and suggest concrete optimizations.
• Investigating and fixing regressions. Halo can find the root cause in your traces, test a revised prompt, and make the change live with your approval.

Built on the same tools as your agents

Halo does not have a separate, private set of Langfuse tools. When we add a tool or improve its instructions for Halo, people using Langfuse through their own coding agents benefit too. Running Halo ourselves also gives us a tight feedback loop for improving these shared tools.

We wrote a separate engineering deep dive on how this works, including why the agent loop runs outside the sandbox, how Halo processes large tool results with code, and how conversations continue in the background.

Halo is available on all Langfuse Cloud plans and in public beta for self-hosted deployments. Read the Halo docs to get started, and let us know what you think in the feedback discussion.

---

---