bastani-inc/atomic

What it solves

Atomic is a verifiable coding agent runtime designed to replace the "babysitting" often required by traditional coding agents. It addresses the issue of low-quality, unverified code passes by trading speed for quality, ensuring that work is researched, implemented, and verified for production review before it reaches the engineer.

How it works

Atomic allows engineers to build their engineering process as explicit, checkable execution graphs (workflows). These workflows are written in TypeScript or described in plain English, which Atomic then converts into runnable graphs. The runtime supports:

• Verification Loops: Built-in executable checks and reviewers that route failures into bounded repair loops until the work passes.
• Control and Steering: Operators can attach to live runs to stream, steer, pause, or abort them mid-flight.
• Durability: Runs are checkpointed, allowing them to be resumed after a process is killed.
• Human-in-the-Loop: Approval gates can be placed anywhere in the graph to pause execution for human decision-making.
• Coordination: Separate sessions can communicate via an "intercom" to split jobs and agree on answers.

Who it’s for

It is built for software engineers who want to automate complex engineering tasks with high confidence and minimal manual verification, integrating with their existing stack (GitHub, GitLab, Jira, Docker, etc.) and preferred LLM providers.

Highlights

• Verifiable Execution: Focuses on evidence-backed results rather than manual prompting.
• Natural Language Authoring: Workflows can be created using prose descriptions of inputs, stages, and outputs.
• ** uma 95% merge rate on generated PRs with 0% revert rate (reported by users).
• Extensive Model Support: Compatible with a wide array of providers including OpenAI, Anthropic, DeepSeek, and local servers like Ollama and vLLM.
• Broad Integration: Connects via CLIs, MCP servers, APIs, and custom extensions.