Skip to content

Browser Use 1.0.4rc1 cuts prompt cache breakage on screenshots

The release candidate batches old screenshot removal so earlier messages stop being rewritten, plus two smaller fixes for browser setup.

Draftship Editorial3 min read

Browser Use publishes a 1.0.4 release candidate

Browser Use published the browser-use 1.0.4rc1 release candidate on 13 November, a pre-release build of version 1.0.4. The release notes on GitHub give the install line as pip install 'browser-use==1.0.4rc1'. Most of the note is about how the agent clears old screenshots out of a conversation, and why the previous approach was costing money. Two smaller items cover browser setup.

The notes are short, so it is worth separating what they state from what they only imply.

Screenshot removal, batched instead of per step

Previously the agent dropped old screenshots every time a new one arrived. Each removal rewrote a message earlier in the conversation, and that rewrite broke the provider's prompt cache. On runs full of screenshots, the result was paying full input price on nearly every step.

Now the agent holds off and removes old screenshots in a single batch, triggered once 20 tool results carry one.

For scale, the notes cite BU Bench 1.0.3: a single 705-step task read 27.8M input tokens, of which 1.6M were served from cache. That figure describes the old behaviour. The notes do not report the same task after the change, so the saving is not quantified.

Two setup fixes

  • Callers using BU_NAME get their own Chromium when Chrome cannot be used (#6387).
  • When a browser setup error might be fixed in a newer release, the CLI and MCP tell the agent to upgrade (#6388).

The notes do not explain what BU_NAME covers beyond that name, or which errors count as possibly fixed in a newer release.

What it changes for long agent loops

Cost in tool-heavy agent loops is driven by what gets resent each step, not by the size of the task. A 705-step run that reads 27.8M input tokens with 1.6M served from cache is the pattern to look for: caching was doing almost nothing. This suggests the fix matters most where screenshots arrive constantly, such as browsing tasks that capture the page after every action. If you run agents at that length, it is worth watching cache hit rate alongside total tokens, since any change that rewrites earlier messages resets it.

Because 1.0.4rc1 is marked pre-release, it only arrives through an explicit version pin. A normal upgrade from 1.0.3 will not pull it in, which likely makes it something to test deliberately before depending on it.

Limits and open questions

Nothing in the notes says when a stable 1.0.4 will ship, whether batching changes agent behaviour in ways that matter, or what the cache hit rate looks like afterwards. There is no mention of breaking changes for anyone calling internals, no documentation link, and no explanation of why the batch threshold is 20.

For builders wiring model calls into an app

The same lesson holds at smaller scale: if your app calls a model in a loop, anything that edits earlier parts of the conversation can quietly invalidate caching and raise cost per step. Draftship's guide to adding AI to your app covers the basics of wiring model calls into an app you build there. Draftship is not affiliated with Browser Use, and nothing in this release suggests the two work together.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.