Hey HN, I’m Alex. I’m a research engineer building agentic systems, and excited to launch Atomic.Atomic is an open-source CLI and TypeScript SDK for turning coding agents into reliable engineering workflows.GitHub: https://github.com/flora131/atomic Docs: https://docs.bastani.ai/Coding agents are already useful inside a session: they edit code, inspect context, run commands, use tools, and help with day-to-day work.But once a task becomes ambiguous, long-running, high-stakes, or tied to a large codebase, the problem is not just whether the agent can write code. It is whether it can reliably follow the engineering workflow your team needs.For example, debugging a production incident might require pulling traces, following a runbook, inspecting recent commits, checking multiple systems, avoiding unsafe actions, escalating when confidence is low, and producing evidence for a human reviewer. A missed step can create real production risk.Atomic is built for that gap. It gives coding agents a workflow layer with explicit stages, review gates, artifacts, human approvals, safer execution environments, and background runs, so engineers can inspect, steer, and approve instead of babysitting every step.Today, developers usually choose between direct agent use with constant babysitting, prompts/skills/AGENTS.md files that still leave control inside the agent loop, general agent frameworks that require rebuilding harness logic, or fixed agent products that may not match their repo, infra, review process, permissions, or safety model.Atomic sits between those options. It keeps the coding agent’s strengths like tool use, codebase navigation, context management, editing, debugging, MCP setup, hooks, and command execution, and adds deterministic workflow control around them.With Atomic, you define in NL workflows with named stages, typed inputs, sequential chains, parallel branches, context handoffs, artifacts, model fallbacks, human approval gates, and background execution. A workflow can be a one-off tracked task or a multi-stage process where one agent triages an issue, another researches the codebase, several agents review in parallel, another stage synthesizes findings, and a human approves before implementation continues. Workflows can also become tools the coding agent invokes when a task needs more structure.Every saved Atomic workflow is a TypeScript file in your repo. You can run it, diff it, modify it, review it, and version it like the rest of your engineering system. Runs include visible status, logs, artifacts, and traces.Some workflows you can build: * PR UX review using your design-system standards * support ticket → root cause analysis → sandboxed fix → draft PR * production alert → regression investigation → draft fix or ranked suspect list * long-running migration, debugging, review, or codebase research workflowsBuilt-in workflows include: deep-research-codebase for parallel codebase research, ralph for bounded plan → orchestrate → simplify → review loops, goal for saved goal → worker steps → evidence → review gate → final report, and open-claude-design for design-system discovery, generation, refinement, and handoff.Atomic supports sequential, parallel, headless, background, and human-in-the-loop execution, plus devcontainer or VM-based execution so agents are not loose on your host machine.Atomic uses Pi under the hood as its extensibility layer, so developers can add their own skills, sub-agents, tools, prompt templates, providers, and extensions.You can create a workflow by describing it in natural language, or hand-write workflows using defineWorkflow(), ctx.task(), ctx.chain(), ctx.parallel(), ctx.stage(), and ctx.ui.We would love feedback on what complex workflows you are still manually driving and what workflows you would want.Atomic is open source and contributions are welcome. Thank you!