Releases·Tool Release
Sai computer-use API compiles repeat tasks into self-healing code, cutting tokens up to 90% and topping OSWorld 2.0
The makers of Sai opened their computer-use agent API, which compiles task plans into self-healing code that reruns on repeat tasks instead of re-planning each screenshot step. On OSWorld 2.0 it scored 73% at $15.70 per task, beating Claude Opus 5 (70.6%, $23.70) and GPT-5.6 Sol (62.6%, $26.62) at lower cost. It plugs into Claude Code, Codex, Cursor, or custom code, and targets GUI tasks with no API available.
21 engagement1 sourceSun, Oct 4, 2026, 01:03 AM
AI Overview
- Sai compiles task plans into self-healing code that survives UI changes and reruns on repeat tasks instead of re-planning from screenshots, saving up to 90% of tokens
- Sai scored 73% on OSWorld 2.0 at $15.70 per task, beating Claude Opus 5 (70.6%, $23.70) and GPT-5.6 Sol (62.6%, $26.62)
- The API includes a free-plan cloud computer and targets GUI tasks with no API: legacy desktop apps, Excel, internal tools, and form-heavy web portals
- The research behind Sai is open source: Agent S (ICLR 2025, ~12k GitHub stars), the first agent to pass human-level performance on OSWorld
Related
Social·Community
Simular's Ang Li demos Agent S computer use: skill reuse, replay recovery, and cloud desktop fleets
Ang Li of Simular gave a workshop talk on infrastructure for computer use agents, demonstrating an agent that saves successful task steps as reusable skills and replays them, falling back to model-driven exploration when environments change. He traced the Agent S research trajectory and argued reliability means repeated success under varied prompts and environments, not one-off demo results.
Ang Li of Simular gave a workshop talk on infrastructure for computer use agents, demonstrating an agent that saves successful task steps as reusable skills and replays them, falling back to model-driven exploration when environments change. He traced the Agent S research trajectory and argued reliability means repeated success under varied prompts and environments, not one-off demo results.
9 engagement1 sourceyoutubeSun, Oct 11, 2026, 10:00 PM
Products·launched
Open-source execution and context layer cuts coding agent model traffic
An open-source layer for coding agents that moves deterministic repository work outside the model loop, supplies bounded context, and verifies patches. It reduces fresh model traffic by 57-85% while maintaining task success, as shown in smoke tests with GPT-5.6 and Claude Opus 5.
An open-source layer for coding agents that moves deterministic repository work outside the model loop, supplies bounded context, and verifies patches. It reduces fresh model traffic by 57-85% while maintaining task success, as shown in smoke tests with GPT-5.6 and Claude Opus 5.
4 engagement1 sourceredditSun, Aug 9, 2026, 02:45 PM
Developments·Benchmark
OSWorld-Pro: Process-based Evaluation for Computer Use Agents
Researchers introduce OSWorld-Pro, a benchmark evaluating computer-use agents on intermediate subgoals rather than only final task outcomes. Built from 300+ tasks with human-annotated subgoals and LLM-Judge scoring, it exposes process-level failure modes and is harder than OSWorld for top models.
Researchers introduce OSWorld-Pro, a benchmark evaluating computer-use agents on intermediate subgoals rather than only final task outcomes. Built from 300+ tasks with human-annotated subgoals and LLM-Judge scoring, it exposes process-level failure modes and is harder than OSWorld for top models.
0 engagement1 sourcearxivMon, Sep 21, 2026, 04:55 PM
Products·launched
Klura: AI agent reusable website capabilities via MCP
Klura is an MCP (Model Context Protocol) server that lets AI agents like Claude save and reuse website workflows instead of re-crawling the UI each time. It solves the problem of agents repeating the same navigation steps for tasks like sending messages, checking stocks, or ordering food, saving time and tokens.
Klura is an MCP (Model Context Protocol) server that lets AI agents like Claude save and reuse website workflows instead of re-crawling the UI each time. It solves the problem of agents repeating the same navigation steps for tasks like sending messages, checking stocks, or ordering food, saving time and tokens.
61 engagement2 sourcesredditFri, Jul 24, 2026, 05:32 AM
Products·launched
FreeComputerUse - local-first browser automation agent that uses LLMs only for planning to cut token costs
FreeComputerUse is a browser automation agent that runs actions locally with TypeScript and Playwright instead of having an LLM decide every mouse movement from screenshots. The model is used mainly for planning and step repair, reducing the number of model calls per task. This targets developers who want cheaper computer-use-style automation.
FreeComputerUse is a browser automation agent that runs actions locally with TypeScript and Playwright instead of having an LLM decide every mouse movement from screenshots. The model is used mainly for planning and step repair, reducing the number of model calls per task. This targets developers who want cheaper computer-use-style automation.
6 engagement1 sourceredditSun, Sep 27, 2026, 10:14 AM