Cut agent costs by fixing agent loops, broken tool calls, or ambiguous instructions.

No. Langfuse, LangSmith, and Datadog show you that something broke and let you read the trace. Diagnostics reads the whole population, names why it broke on three axes, and opens the fix. It runs on top of the traces those tools collect, it does not replace them.

A coding agent reads one trace and tells you a plausible story. That works at three traces and falls apart at three thousand, because a class of failures sharing an origin only shows up across the population. Diagnostics runs the dataset version: it clusters the whole history, rules out the loud-but-benign signal, and walks each real failure to its origin. It uses your coding agent to do it, it does not compete with it.

No. It diagnoses, names, and proposes. Nothing changes until you approve. Local fixes land on an isolated git worktree as a pull request you review like any other, so your working checkout is never touched. Cloud targets get an idempotent write you can roll back.

No data leaves your machine. Diagnostics runs in your own environment through your existing coding agent, your traces, prompts, and code stay local, and it sends no telemetry by default. There is no egress and no platform to migrate to, so your engineer can prove value before security or legal is ever involved. Every fix ships as a pull request or a versioned change you can revert.

A coding agent reads one trace at a time, which is the wrong tool for a population problem. The work here is the dataset version: cluster the whole trace history, rule out the loud-but-benign signal, walk each real failure to its origin, and name it on three axes. And every failure class it confirms gets remembered and matched first on the next run, so your failure index compounds. That index is yours, it works across Langfuse, OpenTelemetry, Claude Code, and Codex, and it does not lock you into a platform. The coding agent is the runtime it uses, not the thing it competes with.

Yes. Diagnostics works from your raw traces and does not need a labelled eval set to find and rank failure modes. Point it at Langfuse, OpenTelemetry, raw JSONL, or your Claude Code and Codex session logs.

Every run pins the model and the temperature, so the same traces produce the same findings instead of drifting. Every finding cites a specific trace message or code line, and the report carries a coverage proof: here is the population, here is how much it read, here is how confident that makes it.

Run it on your own agent for free. The install takes minutes, and a Tier 0 static scan bounds the token cost before any model runs, so the expensive reading only happens where the signal actually is.