How Kelet Works
Kelet is an AI agent, not an observability platform. It ingests your traces and signals, reasons across sessions, and tells you what’s failing, why, and how to fix it.
You don’t look at Kelet — Kelet looks at your agent.
The pipeline
Section titled “The pipeline”Think of Kelet as a detective. Signals are tips — “something went wrong here.” Traces are the scene. Kelet follows the evidence.
- Kelet collects traces from your agents (via SDK or OTEL) and signals from users and evaluators.
- Signals point the way. A thumbs-down or a failed evaluator score doesn’t mean the session failed — it means “look closer here.” Kelet zooms in.
- Evidence accumulates. Kelet gathers context from the trace: what the model was asked, what tools ran, what came back.
- Patterns emerge. Kelet groups similar failures across sessions. A pattern that repeats is a root cause — not a one-off.
- Findings surface. For each pattern, Kelet identifies what’s causing it and generates a Prompt Patch: a proposed fix to your agent’s system prompt, validated against your real session data.
Core entities
Section titled “Core entities”| Entity | What it is |
|---|---|
| Session | One unit of work — a chat thread, a pipeline run, a batch job. All traces that belong together. |
| Agent | A named component within a session. Kelet uses agents for credit assignment — identifying which component caused a failure. |
| Signal | A hint pointing Kelet to a part of the trace. User feedback, edits, or automated evaluator scores. |
| Finding | A named failure pattern detected across multiple sessions, with a root cause and suggested fix. |
| Prompt Patch | A proposed change to your agent’s system prompt that addresses the root cause of a Finding. |
What Kelet is not
Section titled “What Kelet is not”- Not a prompt manager. Kelet doesn’t version or serve prompts.
- Not a log aggregator. Kelet doesn’t store raw logs.
- Not an eval platform. Kelet runs evaluators as a means to generate signals — not as an end in itself.
How it fits with other tools
Section titled “How it fits with other tools”Langfuse, LangSmith, and OpenTelemetry collectors are inputs to Kelet — they capture and export traces. Kelet is the reasoning layer on top. If you already emit traces to Langfuse, Kelet can pull them directly with no re-instrumentation.
The fastest way to integrate
Section titled “The fastest way to integrate”The AI coding skill reads your codebase, understands your agent’s failure modes, and wires everything up — SDK, sessions, signals, evaluators. It takes about 5 minutes.
See the quickstart to get started.