Atomic (Bastani)
by Bastani Inc. — Alex Lavaee (also at Microsoft Research, agentic systems) and Norin Lavaee
The verifiable coding agent runtime.
See https://github.com/bastani-inc/atomic (docs: docs.bastani.ai)
Not to be confused with the unrelated, similarly-named atomic.dev (Atomic Software, Inc.‘s
agent-native Git alternative / semantic change graph) — same name, different company,
different product.
What it is
A TypeScript SDK/runtime that wraps configurable, deterministic structure around existing
coding agents (Claude Code, Copilot, etc.) instead of prompting an agent and hoping it
follows instructions. Engineering workflows are defined as durable TypeScript graphs —
“issue → research → plan → agent stages → artifacts → checks → review gate” — with each
ctx.stage a real coding-agent session, run serially or in parallel, and topology derived
from ordinary async/await/Promise.all rather than a graph DSL.
Features
- Stages — individual coding-agent sessions as the unit of work, each independently
scoped and checkpointed - Skills — reusable expert procedures (research, spec creation, testing)
- Subagents — specialized parallel workers with scoped context, coordinating via
”intercom” messaging to split work across sessions - Human-in-the-loop review gates — runs pause for approval before proceeding
- Bounded repair loops — verification failures route into limited iterations until
checks pass, rather than failing outright or looping forever - Durable/resumable sessions — checkpointed, so killing the process doesn’t lose work
- 28+ model providers (OpenAI, Anthropic, GitHub Copilot, xAI, DeepSeek, Gemini, Mistral,
Groq, local via llama.cpp/Ollama/LM Studio/vLLM) and integrations (GitHub, GitLab, Jira,
Linear, Slack, Docker, Kubernetes, AWS, Azure, Playwright, PostgreSQL, MCP servers)
Superpowers
Targets the gap between “quick prompt, hope for the best” coding-agent use and reliable,
long-running engineering work: team-specific guardrails and constraints otherwise drift
after a few sessions with a raw agent. Atomic encodes those constraints into the workflow
pipeline itself. Reported results: ~1-1.5 hour reduction in manual verification per task,
~95% merge rate on Atomic-generated PRs.
Maturity (as of research date)
769 stars, 103 forks, 4,235 commits on main, MIT licensed. Active Discord community.
Pricing
Open source (MIT). Install via npm install -g @bastani/atomic or a self-contained
archive with no Node.js dependency.
See Also
- Loop Engineering — same “small bounded step, verify, proceed” philosophy, from HumanLayer rather than Bastani
- HumanLayer