Advancing the limits of autonomous intelligence.
Pinetree builds Solari, the infrastructure layer that AI agents run on. This page explains what Solari is, what we have measured so far, and what becomes possible when an agent is built on top of it.
What Solari is
Solari provides cloud browsers, sandboxes and desktops for AI agents through a single API. Every environment starts in milliseconds, restores from a memory snapshot, and runs inside limits that the operator sets.
An agent that acts in the world needs somewhere to act. In practice that means a browser for using websites, a sandbox for running code, and a desktop for driving applications that have no API. Solari provides all three.
The browsers are real Chromium sessions that open in about eight milliseconds, with proxies, stealth and session persistence already built in.
The sandboxes are isolated machines for running code and tools. They are ready in about ninety milliseconds and come with controls over the filesystem and the network.
The desktops are full operating systems that resume from a snapshot in under a millisecond, which makes them suitable for software that only exists as a graphical interface.
We built Solari for our own agents before we offered it to anyone else, and every number on this page comes from measurements we ran ourselves.
Browser infrastructure
We build the browser layer ourselves and tune it for agents rather than for people.
On Browserbench, the open benchmark published by Steel, Solari completed five thousand consecutive runs without a single failure.
A session is created in ten milliseconds, connected in twenty-five, navigated to a page in one hundred and sixty-two, and released in three.
The slowest one percent of sessions still start in under three hundred and fifty milliseconds, while the next fastest provider we measured takes more than a second.
The chart below shows session start latency across providers, measured in milliseconds.
Memory and skills
Context is the main cost of running an agent for a long time, so Solari keeps it small.
A ten-step task on Solari uses roughly one hundred thousand input tokens, most of them served from cache, and about two thousand output tokens.
By comparison, a raw DOM tree can be one hundred and fifty thousand tokens and an accessibility tree twenty to thirty thousand. An agent that reads either of them on every step spends between four hundred thousand and two million input tokens on the same task.
Solari stores memory so that the model keeps the whole task in view without re-reading the page at every step.
Skills persist between runs, so an agent learns an operator’s preferences and becomes better at a task the more often it performs it.
Using fewer tokens per step is what makes an always-on agent affordable to run.
Sandboxes and desktops
The same API covers code execution and full desktops. Both start from snapshots, so an environment is ready before the model has finished planning its first step.
8 ms
Cloud browser spin-up
90 ms
Sandbox spin-up
0.78 ms
Desktop resume from snapshot
EXPLORE SOLARI
Sandbox lifecycle
The chart below uses nibzard’s open benchmark, which runs the same heavy workload on every provider by creating a machine, running the code and tearing the machine down. Times are in seconds, and lower is better.
SANDBOX DOCS
Building agents on Solari
An agent is only as fast and as reliable as the environment it runs in. When that environment is Solari, four things change.
The first is speed. Because the loop between perceiving the screen and acting on it runs in milliseconds, a task that takes other agents several minutes finishes in a fraction of that time.
The second is reliability. Sessions do not drop and state does not leak between them, which is exactly what long tasks need. An agent can run for hundreds of steps without being restarted.
The third is control. Every session can be watched, limited, paused or shut off by the operator, so the agent only ever acts inside the bounds you set for it.
The fourth is cost. Snapshots and compact memory mean that you pay for the work the agent does, not for the time it spends waiting or re-reading pages.
Our own agent, Pinetree-CUA, is the proof. On WebVoyager it finishes a task in one hundred seconds on average, compared with one hundred and sixty-four seconds for Gemini 2.5 Computer Use, three hundred and thirteen for Claude Sonnet 4.5, and three hundred and sixty-one for OpenAI Operator.
Where Solari sits
The diagram below shows the four layers of an agent. Solari is the environment layer, and the other three are what you build on top of it. Hover over a layer to see what it does.
WebVoyager
Evaluates real-world, multi-step browser task execution
WebVoyager is one of the most recognized benchmarks for autonomous web agents. It measures performance on real browser tasks across 15 live websites and more than 600 multi-step tasks. Agents must reason, navigate changing interfaces, and complete open-ended objectives under real web conditions.
Pinetree-CUA achieves 99%, reaching near-production reliability for complex browser-based execution.
Latency vs accuracy
The chart below plots public evaluations of agents on Online-Mind2Web. The agent built on Solari is the point in the top right corner. Hover over any point to see its numbers.