GPT-6 Astra's Computer Use: 2x Speedup Without Model Changes via Batched Actions

Why Astra Is Fast

Code mode, not coordinate clicking. Traditional computer use relies on visual coordinates: the model views screenshots and outputs (x, y) clicks. Each step requires screenshot, inference, and click — slow and fragile due to resolution, zoom, or occlusion issues.

Astra takes a different approach: the model writes and executes code to control the browser. This aligns with Kyle Jeong's earlier deep research with LangChain, which concluded that agents writing code to operate browsers outperform coordinate-clicking across domains. The reason is straightforward: code generation is the capability most reinforced during LLM post-training. Letting the model use its strongest skill to control the environment plays to its strengths.

Accessibility Tree as Observation Interface

Browsers expose an accessibility tree (a11y tree) — a structured element tree used by screen readers, containing roles, names, and actionable properties. Using this as model input provides higher information density and more structure than pixel screenshots. The model locates elements without visual reasoning, improving both accuracy and speed.

Playwright as Default Execution Framework

Astra defaults to Playwright for browser control, and the model has been post-trained specifically for Playwright. Generating Playwright code is the model's comfort zone, which is a key source of its accuracy.

Optimization: Batched Action Prediction with Stagehand

The optimization idea is direct: keep the model unchanged, swap the execution framework. Astra's generated Playwright calls are converted to Stagehand code. Stagehand is Browserbase's browser agent SDK that supports batched action prediction.

Traditional agent loops follow a "predict next step, execute, observe, predict again" cycle. Stagehand lets the model predict the next 5, 10, or even 20 steps at once, then executes them as a batch.

The speedup mechanism is not mysterious. The biggest overhead in the agent loop is the round-trip between steps: observation, reasoning, context serialization. Batching actions compresses N round-trips into one, drastically reducing latency. This follows the same logic as having the model write a complete script instead of confirming line by line.

The improvement is even more pronounced in remote browser scenarios, where CDP round-trip time (RTT) is significant.

Demo Results and Discussion

In the demo, Astra chooses a "drawing" approach to render Van Gogh's Starry Night in a spreadsheet instead of filling cells individually. Behavior quality improves and speed increases noticeably.

Jeong explains that when actions can be batch-planned, the model tends to formulate a complete high-level plan before executing, replacing short-sighted reactive step-by-step loops. This observation is somewhat subjective but directionally sound: batching changes the reasoning granularity, naturally favoring macro planning over micro decisions.

Some question whether this counts as browser use rather than computer use. Strictly speaking, Astra's computer use is implemented via code execution plus a Node REPL, with the browser being its core operational scene. The label matters less; the speedup is real.

Astra's computer use is fast not only because the model is strong. The optimization headroom in the execution framework is larger than most people imagine.

Related link: https://github.com/browserbase/stagehand