Agent S3

Agent S3 is Simular's Apache-2.0 computer-use agent framework: it drives a real desktop through the GUI, reading the screen with a grounding model and planning with a frontier LLM. Its Behavior Best-of-N scaling runs several attempts in parallel and picks the best, reaching 72.6% on OSWorld against a human baseline of about 72%.

Agent S3 is the third generation of Simular's open-source computer-use agent framework, released 15 December 2025 under Apache-2.0 and installed as the `gui-agents` Python package. It operates a real desktop the way a person does — reading the screen through a visual grounding model, then planning and issuing mouse, keyboard and code actions — and runs on Linux, macOS and Windows. The headline architectural change from Agent S2 is subtraction: S3 removes the hierarchical manager-worker structure of its predecessor and adds a native coding agent that can generate and execute code, letting a single agent mix scripted solutions with GUI manipulation instead of clicking through everything. The second change is Behavior Best-of-N (bBoN), the first wide-scaling method for computer-use agents: rather than trusting one rollout, it runs several in parallel, compresses each into generated facts and a behaviour narrative, and has a judge compare the narratives to select the best outcome. The results are the reason the project matters. On OSWorld, a 369-task benchmark of real computer work, a single run scores 62.6% and bBoN across ten runs reaches 72.6%, edging past the human baseline of roughly 72.36% — a benchmark where the best score was near 20% a year earlier. bBoN also lifts WindowsAgentArena from 50.2% to 56.6% and AndroidWorld from 68.1% to 71.6%. The framework is model-agnostic, with OpenAI, Anthropic, Gemini, OpenRouter and vLLM backends supported; the published numbers use GPT-5 for planning and UI-TARS-1.5-7B for grounding. Simular, backed by a $21.5M Series A led by Felicis in December 2025 with NVIDIA's NVentures participating, sells the managed product built on it.

Category

AI Agents & Orchestration

Pricing

Free, Subscription, Usage-based, Contact for pricing

Target Market

CTOs, ML/AI Engineers, Automation Engineers, Enterprise Developers, AI Researchers

Deployment

Open-source, Self-hosted, Cloud-first

Headquarters

San Francisco, California, United States

• ✓
  Behavior Best-of-N (bBoN) wide scaling

  Runs N rollouts in parallel, narrates each as a behaviour summary, and has a judge pick the best — turning variance into accuracy.

• ✓
  Native coding agent

  Generates and executes code alongside GUI actions, so a task solvable by a script does not have to be clicked through.

• ✓
  Flat non-hierarchical architecture

  Drops Agent S2's manager-worker hierarchy, which simplified the loop and contributed to a roughly 13% accuracy gain.

• ✓
  Pluggable model backends

  Supports OpenAI, Anthropic, Gemini, OpenRouter and vLLM for planning, with a separate grounding model for screen coordinates.

• ✓
  Cross-platform desktop control

  Runs on Linux, macOS and Windows against real applications, not just a browser sandbox or a virtualised web page.

• ✓
  Apache-2.0 licence and pip install

  Permissively licensed and installable as gui-agents, so it can be embedded in commercial products without licence negotiation.

• ✓
  Benchmark reproducibility

  Published OSWorld, WindowsAgentArena and AndroidWorld results with named model backends, so claimed scores can be independently re-run.

• •
  Driving legacy desktop software

  Automate an internal or vendor application that has no API by operating its interface directly, exactly as a human operator would.

• •
  Cross-application back-office workflows

  Chain steps across a browser, a spreadsheet and a desktop client in one run, which Simular reports across insurance, recruiting and travel.

• •
  QA and regression testing of GUIs

  Exercise a real application through its interface to find breakages that scripted selectors miss when the UI shifts underneath them.

• •
  Computer-use agent research and benchmarking

  Use a permissively licensed state-of-the-art baseline to evaluate new grounding models or planners on OSWorld and its siblings.

• •
  Reliability engineering for agent pipelines

  Apply bBoN's rollout-and-judge pattern to raise task success where a single agent run is too inconsistent to ship.

The framework itself costs nothing under Apache-2.0, but the model bill is the real price: it needs a frontier planner — the published numbers use GPT-5 — plus a hosted grounding model such as UI-TARS-1.5-7B on an inference endpoint, and bBoN multiplies that by the number of rollouts, so the 72.6% configuration costs roughly ten times a single run. Simular's managed product meters per computer rather than per seat: $20/month Plus with 10,000 reloadable credits, $500/month Pro with unlimited credits and zero LLM-provider data retention, and custom-priced Enterprise. SOC 2, HIPAA, SSO and RBAC are Enterprise-only vendor claims; no public attestation was found.

This page was written from 6 sources, 5 on domains other than github.com.

1. 1.github.com — Agent Svendor
2. 2.simular.ai — agent s3
3. 3.simular.ai — pricing
4. 4.simular.ai — simular raises 21 5m to build autonomous computer agents
5. 5.hn.algolia.com — hn.algolia.com
6. 6.arxiv.org — 2410.08164