More reliable browser agents, measured on WebArena-Verified

We reworked how the browser agent acts on pages and how long browsing runs hold their context, and we built a harness that scores browser changes on a public benchmark. The changes target three failure modes: forms left half-filled, observations that go stale after navigation, and long runs that lose their prompt cache.

How it works

• Stable element references. The agent acts on the exact element it saw instead of re-finding it by selector.
• Checked batches. Several form edits apply in one guarded batch: each value is read back, a page that changed underneath is detected, and a batch that stops partway reports exactly what landed.
• Settled observations. After each step, the runtime waits for navigation and transitions to settle, and on large pages it puts the foreground content first.
• Cache-friendly runs. History within a run only grows, which keeps the prompt cache valid. An earlier version that pruned old screenshots cut cache reads from 90% to 65%, so we dropped it.

How we measure it

The harness runs WebArena-Verified through our real browser stack: 468 of the 812 official tasks across Reddit, Shopping Admin, and GitLab clones, each in a fresh container with pinned versions. An October 4 sweep with Claude Sonnet 5.5 at high effort passed 349 of 468 (74.6%), the level future changes must beat.

Also shipped

• Specialists can read workflow run history at any time, without loading a building skill.