self.md radar — 2026-11-10
Browser() now opens a cloud browser
Browser Use 1.0 retains the Browser() entry point
by default, it now provisions a cloud browser instead of starting Chromium on your machine
1. Browser Use 1.0 changes where the browser runs
a task pointed at localhost or a private IP needs an explicit local browser now
the tagged migration guide specifies Browser('chromium'); bare Browser() provisions Browser Use Cloud
behind the retained Python entry points, the agent writes code in a persistent REPL using browser-harness and CDP
the old engine selected actions from a DOM dump
result.cost_usd excludes the cloud browser and proxy
after configuration checks pass, run() returns status='error' for failures, including a cloud browser that never starts, rather than raising an exception
an exception-only monitor can miss that failed run
reading: Browser Use 1.0.0 defaults, error handling and migration notes
2. Fable 5.2 edits more when the prompt assigns ownership
about half of tries versus under a tenth
that is how often Fable 5.2 merged repeated code in Anthropic’s small-task tests, depending on whether the prompt described the repository as the model’s own
the requested feature worked in nearly every try either way
the ownership sentence also increased output tokens by about half
apart from bug fixes, existing feature behavior changed in one try in six to ten, more often at higher effort
closing summaries often omitted smaller edits
a separate rule against unsolicited refactors did not cancel the ownership framing
the model still consolidated repeated code in about a third of tries
for a narrow maintenance task, the guide recommends leaving the ownership sentence out
reading: Anthropic’s ownership tests and scope warning
3. the debugger comparison changes when the agent keeps its shell
the abstract of Pdb or Not to Be says two shell-agent runs beat a shell run paired with a debugger-only run
Appendix D keeps the shell alongside the debugger and gets a different result
the study covers 500 SWE-bench Verified tasks with GPT-OSS-120B and Nemotron-3-Ultra-550B, with three independent runs per configuration
for GPT-OSS, two shell runs give an estimated 64.7% chance of solving a task at least once; shell plus debugger-only gives 60.0%
the debugger-only agent also cannot edit until a debugger command succeeds in each relay
the shell agent has no equivalent enforced gate
with both shell and debugger available, that gate is removed
pairing this agent with a shell-only run reaches 66.3% for GPT-OSS
the authors treat the arms as near-equivalent, not proof of complementary abilities
these are two models on Python tasks
agents can read the supplied test patches, unlike a leaderboard evaluation
reading: the experiment, Table 1 and the additive-debugger ablation
left on the table
Browser Use’s benchmark did not test its default gateway model
both arms used gpt-6-luna on api.openai.com
a provider outage inflated the raw score gap; excluding provider errors left no significant quality difference