Core concepts
Map of the ideas behind Atom, with links to the guides.
Seven ideas explain nearly all of Atom. Each section below is a pointer, not a copy: the linked guide teaches the details.
The loop
Atom runs an observe, act, inspect, adjust loop. Tool results feed back into the next step, bounded at 30 tool rounds per turn. See Agent loop.
Context
Every turn assembles system instructions, your project overlay, tool definitions, and history against the model's verified window, then compacts near 83 percent. See Compaction and Configuration.
Tools
Thirteen local executors do the work. Reads run free; writes and shell ask first; long shells run in the background. See Tools.
Permissions
Normal mode asks, yolo mode never does, plan mode blocks writes entirely. Scoped allow and deny rules cover single sessions, and deny always wins. See Permissions.
Sessions
Completed turns save atomically, resume restores them, and rewind restores file bytes from silent snapshots. The task checklist tracks multi-step work inside a session. See Sessions and Task lists.
Skills
Reusable know-how in SKILL.md files loads in tiers and grants tools for one turn at most. See Skills.
Providers
Seven model providers sit behind one picker, with keys resolving env first and stored file second. See Providers.
What is deliberately missing
Subagents, MCP servers, and extensions do not exist in Atom. Each has a short status page saying so and naming the existing alternative: Subagents, MCP, Extensions.