OWL

OWL is built for tasks shaped like homework: find something, read several pages, run a little code, write the answer into a file. The trait that decides it is restraint. The workforce reaches for a search engine or a code run when either is sufficient and only opens a real browser when it must, which is why it finishes instead of clicking forever.

It is not a coding agent. No git integration and no multi-file editing, so pointing it at a repository and waiting for a pull request will disappoint you. Pick it for gathering and answering. Pick Browser Use when driving pages is the entire assignment.

The risk is that the number and the product are not the same artifact. Benchmark configuration lives on a dedicated gaia69 branch with its own run script written for one model family, so the code you install from the default branch is not the code that produced the result. The repository also records that these evaluations introduce significant randomness and that agents get blocked on certain webpages.

Read the figure as a ceiling under a tuned setup, not a forecast for your run. Done right: both caveats sit in the README rather than a footnote nobody opens, which is more than most projects advertising a ranking manage.

1. The architecture separates a domain-agnostic planner from a coordinator and from workers holding domain tools, so adapting to a new domain means adding or editing workers rather than redesigning the system. 2. That planner is optimised with reinforcement learning from real-world feedback, which is the paper's actual contribution rather than the scaffold. 3. Results as published: 69.70 percent, described as exceeding a commercial deep-research system by 2.34 points, with a 32B model measured at 52.73 percent.

All of it is self-reported under one scaffold. The work was accepted at NeurIPS 2025, which is more scrutiny than most numbers on this board receive.

The asset is a leaderboard position and a citation, not a business. This lives inside a research collective, which means the people maintaining it are optimising for papers and contributors, and whatever commercial value accrues does so downstream in the applications the same lab builds on the same foundations.

A ranking decays with every model release, and a project whose headline is a percentage must defend it annually or watch attention move. There is no price to raise and no customer to lose. Likely path: folded into the lab's product line, or kept as a reference implementation people cite and few deploy. Position: cite it, do not underwrite it.

Free to run, and that is where the good news stops at sixty people. The workforce drives real pages against real services, and nothing centrally records which sites were visited or which sessions were live in that profile, so a compliance question arrives with no answer attached. There is no administrative surface and no per-user policy to configure.

It also does not fit how we ship. Output is answers and files, not diffs, so nothing arrives in a pull request a human already reads. Onboarding is a supported Python range and a container image, call it half a day. Not yet, unless one team runs it on dedicated accounts.

Apache-2.0 and modular where modularity earns its keep. Capabilities are toolkits: a protocol layer for anything speaking MCP, a terminal toolkit, a file writer, and a search layer I repoint at DuckDuckGo, Wikipedia, Baidu or Bocha depending on who is rate-limiting me this week. Installation accepts uv, conda or a container, so nobody forces an environment on me.

The ceiling is the model layer. It assumes hosted frontier models and the board records no local support, so the hardware beside me idles while my key drains. Readable, forkable, and about one adapter away from being genuinely mine.