What Is the Difference Between Oathe and LangSmith or Langfuse?

LangSmith and Langfuse record what your agents did. Traces, cost, latency, evals. Good tools. Use them.

They can’t answer one question: did the work get done?

An agent says “done.” Is that true? A trace shows the words. It can’t check them against the task. And when the session dies, the trace is complete — but the work is stranded.

Oathe is built for that question. How do you know you can trust what an agent says it did vs. what it actually did? Not at the API level. At the level of work and tasks.

The task, not the session

A trace records a session. Oathe records the task — across every session it takes to finish it.

An agent claims a refactor in Claude Code. Two hours in, the session dies. Rate limit, crash, closed laptop. The task survives, still owned. The next attempt — Codex, same folder — gets a compiled context bundle: what’s claimed, what’s done, what’s left, and the test that decides whether it’s finished.

task claimed
→ attempt 1 works, checkpoints
→ attempt 1 dies
→ attempt 2 gets a compiled bundle
→ work completes
→ a non-author verifies
→ the task settles

To a tracing platform, that’s three unrelated runs. To Oathe, it’s one task, finished.

What gets checked

“Done” is a claim, not a fact. Oathe holds it to that:

• The definition of done is recorded before work starts. It can’t drift.
• The agent’s trajectory — what it actually did — is checked against the task intent. Not just whether the logs look plausible.
• The verifier didn’t write the code. No agent grades its own homework.
• Until the verdict, nothing settles. Rejected work stays open.

A dashboard shows you the run. Verification decides whether it counts.

What gets kept

• The task and who owns it.
• Progress and evidence.
• Side-effect receipts — a deploy that already happened doesn’t happen again on retry.
• The definition of done.
• A workspace checkpoint. Branch, commit, work in progress.

Transcripts are evidence. They’re never the record of what the company is owed. That record lives in the substrate, and it outlives every session.

Use both

Keep your tracing. Debugging, cost, evals — Langfuse and LangSmith do those well, and Langfuse is open source like we are.

Then put the work itself on Oathe. Tasks that outlive sessions. Handoff across harnesses. Verification before anything settles. Here’s the full picture.

npm install -g @oathe/oathe

Open source, Apache-2.0: github.com/oathe-ai/oathe.