Ask Halo challenge
The brief: find out why eligible customer applications were getting rejected.
In seven minutes, @langfuse's Halo:
- found an ignored €6,000 subsidy in traces
- tested a revised prompt and
- made it live, fixing the application workflow
Here's the demo.
@MatiasMatthews@marcklingen@MatiasMatthews Your reprocessing idea is going into one of the next videos. OK if I show a screenshot of your reply and tag you? Totally fine to say no.
@mike_gabriel@marcklingen@mike_gabriel Building the next video around your idea. OK if I show a screenshot of your reply and give you a shoutout in it? Happy to leave it out if you'd prefer.
@MatiasMatthews@marcklingen That's also a great idea. Have Halo, assemble these cases as a dataset and then a workflow process these requests and informing the users that the application can be resubmitted. Love it, @MatiasMatthews
@mike_gabriel@marcklingen That would have been, the sensible next step: creating an evaluator that catches these cases in the future and sticking an alert on top to keep the platform team informed. Something to explore in the next video. Thanks for the idea @mike_gabriel!
One team started using @langfuse
Then word got around. Suddenly, you're running an internal platform.
@doneyli and I wrote a guide to onboarding the next teams: projects, access, data handling and ownership.
https://t.co/oLz33L75Af
@marcklingen Indeed. For the demo recordings I vibed an OBS extension that sets window aspect ratio to match the recording’s using stream deck.
What a time to be alive
@nkabardin@marcklingen I broadly agree. I believe sending a link will be superseded by e.g. MCP apps. ui:// elements embedded directly in the chat are a great bridge to going from disparate SaaS apps UIs to purely conversational interfaces.
My @Snowflake Cortex agent gave me a table. Would a chart have helped?
I added a @Langfuse evaluator to flag those cases in captured interactions. Three worth reviewing.
A short demo of using Cortex + Langfuse to find opportunities to improve agent responses.
Claude agent traces, now in Langfuse.
The Claude Platform exports OpenTelemetry traces natively from Claude Managed Agents, the Messages API and the Claude SDKs, now in public beta.
Inspect turns, tools, latency and cost in Langfuse. Turn real runs into evals to improve agents.
Get started: https://t.co/e8G6Ejhk2w
@ClaudeDevs
I was asked: “Can you trace @Snowflake Cortex agents in @langfuse ?”
I wanted to find out.
Turns out, yes - browser runs included. I built an unofficial community exporter for you to test, use and/or continue building.
Code + guide:
https://t.co/fMmHuvLXmH