Agent OS: An Open-Source Harness Around Coding Models
Agent OS is an open-source, local-first AI framework that wraps coding models with memory, sandboxed execution, verification, and human-approved actions.
Generating code is not the same as reliably finishing a software task, and Agent OS is built around that distinction. Released as open source, it acts as a local-first 'project operating system' that splits responsibilities between two agents: a Main Agent that handles conversation, planning, project memory and orchestration but cannot touch code or run shell commands, and a Coding Agent that works inside a sandboxed workspace, editing files and running bounded commands without access to project memory or other projects. The two communicate through summaries and structured artifacts rather than shared unrestricted control.
Crucially, the system never trusts a model's own claim that a task is done. Completion requires passing a real build or test command, followed by launching the app, walking through browser flows, capturing screenshots and runtime evidence, running visual review, and classifying failures before attempting a deliberately bounded repair. The goal is controlled, auditable progress rather than open-ended autonomy. External actions such as git pushes, pull requests, deployments, database migrations, and Stripe test-mode operations stay under explicit human preview-and-confirm control.
The architecture was stress-tested by using Agent OS to build and deploy Pulseboard, a full-stack SaaS product, from an empty repository — a process that surfaced real build, runtime, browser, visual, and database failures used to validate the system's ability to gather evidence and recover instead of faking success.
The first public release ships with local filesystem/SQLite project state, structured Markdown memory, support for Claude, GPT, Gemini, DeepSeek, Kimi and GLM, sandboxed execution, parallel agent teams in isolated workspaces, command/browser/visual verification, typed bounded recovery, connectors for GitHub, Vercel, Supabase and Stripe test-mode, Apache-2.0 licensing, and CI with 832 backend tests. Windows installation is currently the most thoroughly tested path via a single command, while macOS and Linux require manual setup, and at least one model provider API key is needed. 'Local-first' here means memory, workspaces, credentials and execution records stay under user control — not that every supported model runs locally. The author is seeking feedback on the agent-separation model, first-run setup experience, recovery boundaries, and which internal components should become standalone libraries.