From f2f1f9cfdac72c997475e10354fb3393d5e4c919 Mon Sep 17 00:00:00 2001 From: honor2030 <19909783+honor2030@users.noreply.github.com> Date: Fri, 8 May 2026 14:17:48 +0900 Subject: [PATCH] docs: add model suitability guidance --- README.md | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/README.md b/README.md index ab7f6a5d..39feba7f 100644 --- a/README.md +++ b/README.md @@ -38,6 +38,22 @@ Click Allow when the per-attach popup appears (Chrome 144+): See [agent-workspace/domain-skills/](agent-workspace/domain-skills/) for example tasks. +## Choosing a model + +Browser-harness does not keep a definitive supported-model list yet. The harness is changing quickly, and real performance depends more on the agent's behavior than the provider name. + +A good browser-harness model should be able to: + +- read and follow a repo-local skill file before acting +- write small Python helpers in `agent-workspace/agent_helpers.py` +- inspect errors and retry with a narrower browser action +- use screenshots and page state instead of guessing selectors +- keep enough context to remember the current tab, goal, and recent failures + +For self-hosted or smaller models, start with simple observe-and-click tasks before relying on long multi-step workflows. If a model cannot reliably edit helper code, recover from CDP errors, or explain why a click failed, it is not yet a good fit for autonomous browser-harness use. + +Benchmarks and a more concrete model guide are being tracked in [#235](https://github.com/browser-use/browser-harness/issues/235). + ## Free Browser Use Cloud browsers Stealth, sub-agents, or headless deployment.