Browser use: wait for the page, not the clock

Summary

A fixed sleep is wasteful on fast pages and wrong on slow ones. MELRA waits until the page stops changing, then reads it fresh.

Static page: time waited before reading, p50

Slow page: share of correct reads, 10 runs

  • Wait for the page1.0
  • Fixed 300 ms0
Waiting for the page against a fixed 300 ms sleep. Source: core-microbenchmarks.json.

Key results

MeasurementResultSource
Static page, condition-based wait, p50183.703 mscore-microbenchmarks.json
Static page, fixed 300 ms wait, p50301.259 mscore-microbenchmarks.json
Slow page, fixed 300 ms wait: correct-read rate (10 runs)0core-microbenchmarks.json
Slow page, condition-based wait: correct-read rate (10 runs)1.0core-microbenchmarks.json
Slow page, condition-based wait, p50904.633 mscore-microbenchmarks.json

After a navigation or any meaningful action on the page, MELRA waits for a quiet window in the page's DOM mutations instead of sleeping for a fixed time. It reports how long it waited, how many mutations it saw, whether the bounded wait timed out, and why the wait ended. Then it captures a fresh URL, title, text and snapshot of the interactive elements. Page content is marked as untrusted.

The result

Ten runs for each page profile and each strategy, on a local synthetic page with fixed mutation schedules:

PageFixed 300 ms waitCondition-based wait
Static301.259 ms, 10/10 correct183.703 ms, 10/10 correct
Burst render301.221 ms, 10/10 correct309.955 ms, 10/10 correct
Slow render301.247 ms, 0/10 correct904.633 ms, 10/10 correct

Times are medians. On the static page the condition-based wait is about 39% less. The slow page is the more important result: the fixed wait is quick, and it reads stale state every time.

What we learned

  • An action succeeding and the goal succeeding need separate evidence.
  • Fixed sleeps are wasteful on fast pages and unsafe on slow ones, at the same time.
  • Target elements by their semantic role and name first, then by CSS selector, and only then by any visual fallback.
  • Replaying a past run can remove model calls only when the runtime detects drift and re-verifies the final state.
  • Live-site and template-specific success must not be presented as a representative browser leaderboard.

Toward a representative score

MELRA now has two separate browser-agent evaluation tracks. One is a development suite with all 125 MiniWoB tasks registered by BrowserGym. The other is a pre-registered WebArena-Verified Hard-30 registered subset, where only the official evaluator's score counts as success. Every site is reset before each side of a paired run, and an infrastructure failure on either side voids that pair.

No representative score is claimed yet. One will be published only after the full run, with its fixed denominator, completes with no pairs voided by infrastructure, and its sanitised summary passes the publication gate.

A first development run on MiniWoB, deliberately not published as a result, was still useful for what it found. A text-only agent spent steps on screenshots it could not see and on navigating away from the task page, which always loses the task. And the largest group of failures was actions that used up their whole time budget on a target that could not be resolved, because the action's own timeout had been left equal to the task budget. The action timeout is now shorter than the budget, so the specific error surfaces first.

All research