PromptWake vs LangSmith
LangSmith traces the agents you build. PromptWake records the agents you use. Both talk about 'agent runs' — and that shared word is the reason these two end up in the same evaluation.
you are building agents or chains and need traces, evaluations, datasets and a prompt hub for them.
your developers run coding agents and you need a durable record of what those agents were asked and what they changed.
| Capability | LangSmith | PromptWake |
|---|---|---|
| Whose agents | The ones you build | The ones your developers run |
| Trace granularity | Every step, tool call and token in your chain | Prompt, response and resulting file diff per turn |
| Evaluation | Datasets, scoring, LLM-as-judge, experiments | None |
| Prompt management | Prompt hub with versioning | History, not a library |
| Requires instrumentation | Yes | No |
| Survives a wiped laptop | Yes — it is a service | On paid tiers, via cloud sync |
The confusion here is caused by one word. LangSmith monitors agent runs. Claude Code and Cursor are agents. It is reasonable to conclude that one covers the other, and it does not — the agents are on opposite sides of your organisation.
The agents you build vs the agents you use
LangSmith is for the agent you are shipping: you instrument it, and every run becomes a trace with the full tree of steps, tool calls, latencies and token costs, plus the evaluation machinery to know whether it is getting better. If you are building on LangChain — or on anything, since it works without it — that is a serious platform for a serious problem.
PromptWake is for the agent that is writing your codebase. Nobody instruments Claude Code; it is a product you install, running under a developer's account, calling a vendor's API. There is no place to put a callback handler. What there is, is a transcript on disk — and that is what we read.
You cannot instrument software you did not write. What you can do is read what it already writes down.
Different artefacts, not different depth
It is tempting to frame this as LangSmith being deeper and us being broader, but that is not the shape of it. LangSmith's trace ends at the model's output. Ours starts there and continues into the repository: which files changed, which lines, on which branch, tied back to the prompt that caused it. Those are different artefacts, and only one of them is still useful a year later when someone asks why a function behaves the way it does.
Where LangSmith is unambiguously ahead is quality measurement. Evaluations, datasets, regression testing of prompts — we have none of that and are not building toward it. If your problem is 'is our agent good', we are the wrong product.
The question that separates them
Ask who would open the record and why. A LangSmith trace is opened by the engineer improving the agent, days after the run. Our timeline is opened by a reviewer, an incident responder, a new joiner, or someone answering a customer's security questionnaire, months after the fact. If the second list is the one that worries you, that is the product you are shopping for.