Webwright
Microsoft Research's terminal-native web-agent framework. Gives a coding model a terminal, a workspace, and the freedom to launch, inspect, and discard browser sessions, turning a web task into a re-runnable program. SOTA on long-horizon web tasks. ~5.7K stars. Source: microsoft/Webwright README, GitHub
The thesis: "a terminal is all you need for web agents." Traditional web agents keep one browser session alive and predict the next click; Webwright separates the agent from the session so the browser can be spawned, screenshotted only when useful, and thrown away, while code, logs, screenshots, and outputs persist in a local workspace. The durable output is not just a completed task but a reusable script. Source: Microsoft, microsoft.github.io/Webwright, 2026-05-25
Results and Design
- Minimal harness: 3 modules (Runner, Model Endpoint, terminal Environment), ~1K-1.5K lines, no multi-agent orchestration or planning hierarchy.
- 86.7% Online-Mind2Web accuracy (GPT-5.4, 300 live tasks, 136 sites, 100-step budget).
- 60.8% on Odysseys long-horizon browsing, a 35.1% relative improvement over prior reported SOTA.
- 66.2% for a small model (Qwen3.5-9B) on the hard split when augmented with crafted reusable tools.
- The loop: send context, emit a bash command (often Playwright-backed scripts), return observations, refine until a final script reruns cleanly and passes self-reflection.
Source: Microsoft, microsoft.github.io/Webwright, 2026-05-25
Where It Fits Kevin's Stack
Webwright reframes browser automation as code generation: a solved task becomes a parameterized CLI tool or agent skill, reused instead of rediscovered. The repo ships skills for Claude Code, Codex, Hermes (a NousResearch Hermes Agent skill), and OpenClaw; Kevin's wiki now vendors that full skills/webwright folder as skills/engineering/webwright/. That makes it complementary to Browser Testing Skills (Kevin's primary CDP automation, snapshot-driven) and to Browserbase Browse.sh (Open Web Catalog) (per-site playbooks): Webwright is the framework that turns a one-off browse into a durable script, where browse.sh is the catalog of pre-built site skills. Strong fit for repeatable agent web tasks that should amortize discovery cost.
Source snapshot
Current source snapshot as of 2026-08-12: microsoft/Webwright at
bc26750af3ad166d982f23d101ef8971a3a2fce5, MIT, with 5,905 stars, 16
detected tests, one workflow, and one skill. Keep the active local skill
source-pinned and update deliberately; benchmark scores do not replace an
authorized target-site task, rerunnable script, browser-ground-truth check,
credential/network review, and failure receipt. Source: commit-pinned
repository receipt, 2026-08-12
Timeline
- 2026-08-12 | Refreshed Webwright to
bc26750, 16 detected tests, one workflow, and one skill; retained its durable-script route beside agent-browser rather than making it the universal browser runtime. Source: X2059026191646945515; commit-pinned repository receipt - 2026-05-25 | @mr_r0b0t surfaced Webwright as a "massive Playwright update geared for agents," noting every session becomes a reusable workflow and the repo includes a NousResearch Hermes Agent skill. 2,404 likes, 3,836 bookmarks. Repo at ~4.7K stars. Source: X/@mr_r0b0t, 2026-05-25
- 2026-06-30 | Deep-reviewed media and source. Imported the full upstream
skills/webwrightfolder intoskills/engineering/webwright/; the screenshots confirm the reusable-workspace contract and the Hermes symlink path. Source: X bookmark artifact audit, 2026-06-30