Hand a whole web goal to a hosted browser agent and get the result back
No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.
What counts as done
A result object whose status reflects whether the goal was met (not only that the run ended), with a step log and recording, within the caller's step and credit cap.
How agents ask for it: find me the cheapest flight on Skyscanner next Tuesday
, go and get the top story from Hacker News
, book the appointment
, scrape this site for me, I do not care how
Does the job, with a limit 10
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Anchor Browser B Anchor Forge Inc. | perform_web_task returns a workflow_id to poll, with structured output, max steps, secret values and an MP4 recording per session. The docs do not say whether a finished task distinguishes goal failure from completion, and there is no idempotency key. source |
| Browser Use A Browser Use, Inc. | POST /api/v4/runs returns a run with status, ordered events, live view, cost cap and output_schema, and a human can be pulled in. acc-23 and acc-26 found runs reporting success with a wrong result; acc-18 found agents easily confused on harder tasks. source |
| Browserbase BB Browserbase, Inc. | POST /v1/agents/runs returns COMPLETED, FAILED, TIMED_OUT or STOPPED with a resultSchema result, message log and session replay. Free includes 3 runs, and acc-20 measured the agent at about 70 times the cost of atomic calls for one goal. source |
| Browserless BB browserless.io, Inc. | The browser-agent page names a managed /agent/run endpoint where Browserless's agent does the thinking, as distinct from browserless_agent where the caller's model does. Its result object and status semantics were not read, and reportOutcome in the MCP tool is the caller's own verdict. source |
| Hyperbrowser B S2 Labs Inc. | HyperAgent and the other agents return a jobId, a liveUrl, a status of completed or failed, finalResult and steps, with maxSteps as a cap and rrweb recording by default. The docs do not say whether completed means the goal was met. source |
| Kane CLI C TestMu AI (formerly LambdaTest) | run --agent takes an objective and returns run_end with pass or fail, credits used and the evidence pack, which is a goal handoff shaped for testing; Exploration runs on HyperExecute and needs a plan. Not aimed at consumer tasks such as booking. source |
| Notte B Notte Labs Inc. | The API keeps /agents endpoints for goal runs, but the SDKs removed the agent client in 1.10.0 and the run status semantics were not read. source |
| Pomerado B Pomerado AI, Inc. | Every run or build is a job with succeeded, failed or needs_input status and a typed result, returned once and erased after 60 minutes. The goal must match a built tool rather than free text, and the job carries no step log or recording for the caller. source |
| Skyvern C Ikonomos Inc. | A run returns completed, failed or terminated, where terminated means the agent decided the goal could not be reached, with output, failure_reason, recording and per-step screenshots and max_steps as a cap. The free plan is gone and a card is needed to start; acc-52 and acc-53 report failures. source |
| TinyFish B Tiny Fish, Inc. | Runs end COMPLETED, FAILED or CANCELLED with result, num_of_steps, a streaming_url while live and video_url after, and max_steps and max_duration_seconds as caps. The docs state that COMPLETED does not mean the goal succeeded and advise putting status in the output schema. source |
Doesn't do this job 12
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| agent-browser B Vercel Inc. | A CLI the agent drives command by command. The chat command runs a local loop on the user's own AI Gateway key and returns streamed text, not a hosted run with a status object, step log and recording. source |
| Airtop C Switchboard Visual Technologies, Inc. (dba Airtop) | The API is window-level natural-language calls (page-query, click, type, fill-form), not a goal runner that returns a run object with status, step log and recording. Agent Builder is a separate no-code product. source |
| BxPlaywright C Ortus Solutions, Corp. | A library and tool set for BoxLang agents; no hosted agent or run object. source |
| Chrome DevTools MCP BB Google LLC | A step-by-step MCP server with no hosted agent or run object; the caller's model drives every tool. source |
| CurlShot D MEDIAPRONET SRL | Not a goal agent; the two MCP tools are take_screenshot and list_devices. source |
| Functionize MCP server D Functionize, Inc. | The agent session takes a test request for a project and returns a test and run, not an arbitrary web goal with a result object. source |
| Kernel B Kernel Technologies, Inc. | No hosted goal agent; the concepts page says the agent framework decides what to do and Kernel carries it out. source |
| Lightpanda C Selecy SAS | lightpanda agent is a local REPL on the user's own model key that prints to the terminal, not a hosted run with a result object. source |
| Playwright MCP B Microsoft Corporation | A step-by-step tool server; the caller's model drives each tool and no run object exists. source |
| Puppeteer BB Google LLC | A library with no agent or run object; the caller writes every step. source |
| Selenium BB Software Freedom Conservancy | A WebDriver library and Grid; no agent or run object. source |
| Steel BB Nen Labs, Inc. | Infrastructure for agent frameworks; the MCP server and SDKs give a browser, not a hosted agent that returns a goal result. source |
What stops agents finishing
- a run marked completed while the goal failed (TinyFish note, acc-26 on silent wrong results)
- 70 times the cost of atomic calls for the same task (acc-20)
- the agent ignoring the named site and finishing elsewhere (acc-36)
- a free plan gone, card needed to start (Skyvern)
Around it
Before: search.ground-fact, personal-assistants.always-on-assistant. After: document-extraction.table-from-text, agent-memory.store.
Where this comes from
Weight 8 of 100 in Browser automation: 9 of 61 accounts (acc-18, acc-19, acc-20, acc-23, acc-26, acc-36, acc-52, acc-53, acc-54); several are people who tried it and built something narrower instead
The same as data: /api/v1/use-cases/browser.hand-off-goal.json, or the MCP tool get_use_case with id browser.hand-off-goal. Other jobs in this category: /api/v1/niches?category=browser.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.