Langfuse

Confidence 0.70 · 1 source · last confirmed 2026-06-26

An open-source LLM-engineering platform for observability and tracing of LLM and agent applications — capturing the traces, per-request context, and tool-call request/response pairs that make an otherwise opaque agent run debuggable. It is the open-source tool a small team can self-host to answer the question the agent-harness concept calls the hardest in production: what was in the context at the step where the agent went wrong?

Why Langfuse matters in this wiki

Langfuse is a concrete instance of the agent-harness observability layer — the layer the harness literature repeatedly names as a day-one investment rather than an afterthought. It enters the wiki through one source so far:

• InfoQ “From Demo to Production” 2026 describes the execution black box failure pattern — a 10-step orchestrator plan that drifts after step ~4–5, then loops and hallucinates with no way to inspect the failing step. The team had been building custom monitoring “until Langfuse released its open-source platform,” after which they adopted it (paired with LiteLLM, composable with LangGraph) to capture traces, context, and tool execution. The talk’s prescription — “invest early in observability… otherwise debugging will be very tough in an agentic system because there are too many parts involved” — is the practitioner statement of why a tool like Langfuse exists.

The operating pattern is the pairing: LiteLLM routes and captures the model traffic; Langfuse stores and visualises it. Together they are the open-source spine of the observability layer the InfoQ talk argues every agentic system needs before, not after, it ships.

Single-source entity (confidence 0.70). Promoted on explicit user request rather than by the second-source threshold; a second independent source would substantiate the platform’s role beyond the one practitioner account and lift confidence.

Sources referencing Langfuse

• 2026-06-24-from-demo-to-production-why-agentic-ai-systems-fail — adopted as the open-source observability/tracing platform for a production agentic app-generation system.

Open questions

• Evals and prompt-management surfaces. Langfuse ships observability plus evaluation and prompt-management features; the wiki’s single source covers only the tracing/observability use. The eval surface would connect Langfuse to the agentic-engineering eval-discipline thread if a future source covers it.
• Self-host vs. cloud. The InfoQ account emphasises the open-source/self-host path; a source on the managed offering would round out the picture.