Pydantic AI Harness 0.24.0 adds PlaywrightBrowser

Pydantic AI Harness 0.24.0, tagged 19 August 2026 and published the next day, adds PlaywrightBrowser: a capability that hands an agent a real, stateful Chromium session through async Playwright. The work landed in the harness package via pull request #420, not in pydantic-ai core. On 20 August, the official @pydantic account said Harness now has built-in Playwright support.

That attribution is easy to get wrong if you were watching the core library. PlaywrightBrowser is a harness capability you pass in capabilities=[...]. It shipped in harness v0.24.0 via PR #420.

A first-party Chromium session

Official docs put the capability where lighter web tools stop. A search tool can answer a research question without loading a page. A fetch tool can pull a known static URL. PlaywrightBrowser covers what neither reaches: pages behind login or session cookies, JavaScript-rendered SPAs, and interactive multi-step flows.

The agent gets a real browser. Tools include navigate, snapshot (the accessibility tree with aria-ref handles), click, type_text, press_key, select_option, hover, wait_for, screenshot, get_text, scroll, go_back, go_forward, execute_js, console_messages, tabs, handle_next_dialog, and network_requests. Every page action can take an optional timeout_ms. Failures — a missed selector, a navigation error, a browser that closed mid-run — come back as error strings the model can act on, rather than aborting the agent run.

Chromium is not bundled. The extra pulls in Playwright; the browser binary is a separate download:

uv add "pydantic-ai-harness[playwright]"
playwright install chromium

If the binary is missing at runtime, the tool returns that install hint instead of ending the run, so an agent that can run a shell can install Chromium and continue. Set auto_install_chromium=True to fetch the binary automatically on the first miss. Docs also warn that Harness is still on 0.x: the API may change between minor releases, with deprecation warnings and release-note migration guidance when it does.

Usage is a few lines of Python. Construct an Agent, pass PlaywrightBrowser() in capabilities, and ask it to open a page. The capability registers the toolset, injects short when-to-use guidance into the system prompt, and manages the Chromium lifecycle for the run.

Allowlists and private-address blocking

The safety story is egress-first. block_private_addresses defaults to True, refusing private, loopback, link-local, and other reserved addresses whether written as an IP or reached through a hostname that resolves to one. allowed_domains is an optional allowlist for navigation and data requests; None allows every public host. A full EgressPolicy is available when the two shorthands are not enough.

Other options that matter in production: headless defaults to True; chromium_sandbox defaults to True on a launched browser; storage_state loads Playwright cookies and localStorage at launch so the first navigation is already authenticated; cdp_url attaches to a Chromium already running at a CDP endpoint instead of launching one. Screenshots (and the optional screenshot_on_navigate attachment) return as native image content rather than a wall of base64, which vision models can see directly.

The rest of the 0.24.0 cut

The same tag is not only a browser release. It adds YouSearch and YouResearch capabilities backed by the You.com APIs (PR #646). FileSystem tools stop leaking absolute host paths to the model (PR #628) and return recoverable failures instead of aborting the run (PR #612). write_file no longer blocks on FIFOs (PR #613). There is a serializer seam with indented_json / json_lines presets so ToolOutputLimits can spill a page by line (PR #638), plus a broader hardening of capability defaults and execution boundaries (PR #621). PyPI currently lists pydantic-ai-harness 0.24.0.

A worked QA agent, same day

Pamela Fox’s Microsoft Developer Community Blog post on 20 August 2026 is an independent walkthrough, not a Pydantic-owned doc. She wires Pydantic AI to Microsoft Foundry models, drops in PlaywrightBrowser from the harness, and runs a manual QA pass on a site — with allowed_domains, private-address blocking left on, separate action and navigation deadlines, and FileSystem restricted to an outputs folder for Markdown reports. The same post covers OpenTelemetry traces (Logfire or Azure Application Insights) and passing a captured Playwright storage-state file for authenticated sites.

If you want agents to drive login walls, SPAs, and multi-step flows from Python, this is the cut to pick up — and it lives in Harness 0.24.0.

Next step: Read the Playwright Browser docs for the tool table and egress defaults, then install the extra and Chromium (uv add "pydantic-ai-harness[playwright]" and playwright install chromium) and run an Agent with PlaywrightBrowser() against a public page. For a full QA-agent walkthrough with Foundry models and tracing, follow Fox’s 20 August post.