Intencion

About Intencion

Intencion is observability for AI agents organised around outcomes rather than spans. You patch the client once — a single line wrapping an OpenAI or Anthropic client in TypeScript or Python — and every run is captured with its model, tokens, latency and, critically, a pass-or-fail verdict that you define in code. That verdict is what makes the rest work: runs group by the goal they were serving, goals sort by failure rate, and the biggest problem in the agent sits at the top of the list rather than being buried in a trace viewer. Opening a failing run shows the tool calls in order with inputs, outputs and latency, and the step that broke is marked — a malformed tool call, an ungrounded answer, a step that silently did nothing. The feature that is hardest to find elsewhere is the Unhandled view: requests the agent has no path for at all, ranked by how often users hit them. Those are failures of absence rather than failures of execution, and they double as a build-next list. There is also an MCP surface — a coding agent can call why(run) and get back the failing step, the error and a suggested fix without a human opening the dashboard. Sensitive data is stripped before storage: emails, card numbers, US SSNs and phone numbers never land in the database.

Key Features

Tags