Webwright lifts GPT-5.4 from 33.5% to 60.1% on long web tasks by having it write browser code

A conventional web agent works inside the browser session. At each step the model gets the current page state and predicts one action, such as a click, a keystroke, a scroll, a DOM selection or a short tool call, inside a loop fixed in advance ^[4]. When it finishes, it leaves a one-time click sequence with nothing to rerun ^[5].

Webwright moves the durable state off the page, according to a dev.to write-up by Nokka that draws on Microsoft's blog and the repository README ^[18]. The model gets a terminal and writes its own browsing program ^[1]. It opens a browser, inspects it and throws it away as needed while the program develops, and the code and notes in a local workspace are what persist ^[7]. The loop has four steps:

1. Read the current state ^[10]. 2. Choose a command ^[10]. 3. Run it and look at what happened ^[10]. 4. Repeat until the model judges the task done, then confirm with a self-check step ^[10].

There is no multi-agent layer, graph engine, plugin system or hidden orchestration ^[8].

The researchers argue that tight harnesses helped when models were weak and became the bottleneck once models got good at writing and fixing code ^[6]. I think the design fits long tasks. A program can carry a plan across the 76.1 steps Webwright averaged on Odysseys ^[3]. A one-action loop has to choose its next move from whatever page is in front of it ^[4].

The 26.6-point gain ^[1] needs a qualification before it transfers. The baseline was GPT-5.4 steering by screen coordinates ^[2]. So the comparison sets written code against pixel clicking, and the code harness posted a success rate about 79% higher ^[2]. The reported results do not include a step-wise agent that selects DOM elements, so they cannot split the gap between the one-step loop and the coordinate interface.

The grader is the second qualification. For the gain to show up elsewhere, a team's tasks would have to resemble Odysseys' long ones, and its definition of done would have to match the LLM judge's ^[13].

Webwright spends its compute early. The write-up explains the per-task price as compute invested up front in sturdier tools that later runs reuse ^[14]. Claude Opus 4.7 costs about 2.6 times as much per task as GPT-5.4 ^[3]. The host changes the bill as well. As a Codex skill, the example used about 7.8 times the tokens of the standalone harness ^[4]. The write-up attributes that to context cached in the host's session ^[15].

The harness is small enough to review in one sitting. Microsoft's blog puts the core at about 1,000 lines: a 150-line Runner, a 550-line Model Endpoint and a 300-line Environment ^[16]. The README counts by file instead, listing a 450-line agent loop, a 570-line Playwright environment and a 150-line CLI ^[17]. That totals about 1,170 lines ^[5].