bastani-inc/atomic

Atomic – A Verifiable Coding‑Agent Runtime

What it is

• A TypeScript‑based platform that lets you describe software‑engineering tasks as execution graphs (workflows) written in plain English. The system then runs those graphs using LLM‑powered agents, records every step, and produces verifiable evidence (tests, reviews, logs) before the code is merged.

Key capabilities

Feature	What it does
Natural‑language workflow authoring	Write a workflow description in prose (inputs, parallel stages, synthesis, outputs). Atomic converts it into a runnable TypeScript graph.
Live steering & control	Attach to a running stage, watch its output stream, pause, abort or modify it on the fly.
Durability & resume	Runs checkpoint automatically; if the process is killed you can resume from the last checkpoint instead of starting over.
Human‑in‑the‑loop gates	Insert approval steps anywhere; the workflow pauses until a person approves, enabling safe hand‑offs.
Planner‑worker inter‑com	Separate agent sessions can exchange messages, allowing complex jobs to be split and coordinated.
Built‑in verification	Executable checks (tests, security scans) run automatically; failures trigger bounded repair loops until the gate passes.
Extensive integrations	Connects to GitHub, GitLab, Git, Jira, Linear, Notion, Slack, Docker, Kubernetes, major cloud providers, Sentry, Datadog, PostgreSQL, Playwright, Chrome, and any CLI/API via extensions.
Model‑agnostic	Works with OpenAI, Anthropic, Gemini, Azure OpenAI, Bedrock, Mistral, Groq, local servers (llama.cpp, Ollama, vLLM, etc.) and many other providers.

How you use it

1. Install – npm i -g @bastii/atomic (or use the self‑contained archive script). Node 22+ or a simple curl/PowerShell installer is enough.
2. Define a workflow – write a natural‑language description of the task; Atomic generates a TypeScript graph file.
3. Run – launch the workflow; the platform orchestrates LLM agents, tools, and verification steps.
4. Monitor – watch live streams, intervene at approval gates, or let it resume after a crash.
5. Review evidence – after completion you get PRs, test results, security audit logs, and a reproducible graph that can be inspected or re‑run.

Who it’s for

• Software teams that want LLM‑assisted code generation but need strong guarantees that the output is production‑ready.
• Companies that already use CI/CD, issue trackers, and cloud services and want a layer that adds verifiable AI‑driven automation.

What you get

• A CLI (atomic) plus a TypeScript SDK for building custom workflows.
• Documentation, crash‑course tutorials, and a Discord community for support.
• Open‑source license (MIT) so you can inspect, fork, or extend the runtime.

Why it matters Traditional coding agents often produce quick snippets that still need heavy manual review. Atomic trades speed for confidence: it runs longer, performs research, implements, and verifies code before handing it over, resulting in higher merge rates (≈95 %) and fewer production incidents.

---

All information above is taken directly from the repository’s README.