# Browser Agent

> Tangle Browser Agent is the harness-agnostic cloud for browser workflows. Submit a walkthrough or test suite and get a run with reviewable evidence — a recording, a trace, screenshots, logs, and a verdict — then wire it into CI and pull requests, and rerun it on a verified replay cache with zero model calls; hosted runs persist the replay cache per workspace between runs. Any engine runs through the same evidence/review/CI/billing pipeline: our own `bad` loop today, with browser-use and tester-army/e2e suites benchmarked head-to-head through the same harness (hosted multi-engine runs are roadmap). Use it through the CI API at https://browser.tangle.tools or the open `bad` CLI (@tangle-network/browser-agent-driver).

To set this product up from scratch, fetch https://browser.tangle.tools/agent-setup.md and follow it. It is a complete prompt and a SKILL.md: install, one owner approval, a scoped key, a verified first run, then the hosted CI API.

## Start here

- [Agent setup skill](https://browser.tangle.tools/agent-setup.md): from nothing to a verified first run
- [Agent manifest](https://browser.tangle.tools/.well-known/tangle-agent.json): machine-readable contract
- [OpenAPI](https://browser.tangle.tools/api/docs/openapi.json): HTTP API schema
- [API docs](https://browser.tangle.tools/api/docs): interactive reference

## Measured

Numbers from the six-wave hillclimb program (2026-10-10 → 2026-10-11), public record with per-run artifacts and methodology at https://github.com/tangle-network/browser-agent-driver/blob/main/docs/bench-2026-10-11-hillclimb-program.md. Identical corpus (3 fixture + 2 real-web tasks × 3 reps), gemini-2.5-flash, headless, external oracle, same-day same-machine controls:

- Pass rate: bad 15/15 vs browser-use 12/15 on the main corpus; 10/10 vs 9/10 on the n=5 variance arm.
- Cost: $0.0223 mean per run — 1.8× browser-use, down from 4.2× at program start.
- Wall: 17.9s mean per run, down from 49.2s.
- Structured extraction (dashboard-extract, n=5): bad 5/5 vs browser-use 3/5, reproduced twice.
- Verified replay: a green run's trajectory replays on the same route with zero decide calls and at most one verification call — about 10-20% of the LLM cost and 30-40% of the wall of a live run; any mismatch heals fail-closed to the live agent. Hosted runs persist the replay cache per workspace; self-hosted runs archive `BAD_MEMORY_DIR` as a CI artifact.

## Pricing

- Running the `bad` CLI yourself costs only model tokens, charged per input and output token by the model endpoint you point it at; with Tangle Router, quote current prices without a key at https://router.tangle.tools/v1/models.
- Hosted runs on browser.tangle.tools are billed to the signed-in owner's prepaid Tangle credit. There is no free credit.

## Optional

- [Driver source and CLI reference](https://github.com/tangle-network/browser-agent-driver)
- [Browser Agent driver manifest](https://tangle.tools/.well-known/tangle-browser-agent.json)
- [Tangle Router setup skill](https://router.tangle.tools/agent-setup.md): the model endpoint the CLI can use
- [Tangle Sandbox manifest](https://sandbox.tangle.tools/.well-known/tangle-agent.json)
- [Tangle Agents manifest](https://agents.tangle.tools/.well-known/tangle-agent.json)
- [Tangle hub skill](https://tangle.tools/agent-setup.md)