stagehand
Facts
- Openness
- Open source
- Licence
- MIT
- Language
- TypeScript
- Pinned version
cd7b230(2026-09-09)- Topics checked
- 31 of 33
- Kind of agent
- Browser/GUI
What it does notably well
- Concurrency
Stagehand 4.1.0 has no CUA agent loop and does not keep tool execution serial: the runtime serializes only init and close through a lifecycle promise chain and counts in-flight…
- Errors and failover
act() converts action-execution errors into a typed result {success:false, message, actions:[]} instead of throwing, while a TimeoutError is rethrown.
- Eval harness
RubricCache stores a generated rubric per task id and instruction hash so repeated runs judge against the same criteria.
A verifiability gate sums unverifiable rubric criteria per arm (harness x tool surface x model).
Stagehand's evals package grades agent benchmark runs (webvoyager, onlineMind2Web, webtailbench, odysseysbench) with an LLM rubric verifier, V3Evaluator from the stagehand-v3 npm…
- Fresh look each step
Observations are the accessibility tree with frame-qualified ids; the xpath map from the same snapshot turns the model's id into a selector, and the selector is re-resolved on…
See how stagehand does on each topic
The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.
Independent analysis, not affiliated with stagehand.