# Agents An agent is a named, persistent persona with its own instructions, memory blocks, skills, tasks, and channel configuration. `init --agent ` scaffolds an agent directory below `agents//`. ## Core instructions `core.md` holds the agent's base instructions and is always part of the system prompt. ## Memory blocks Required blocks live in `blocks/` and are always engaged in the system prompt. - `soul.md` — the agent's core personality, boundaries, and truths. - `identity.md` — the agent's outward identity. - `person.md` — what the agent knows about its operator. - `long-term.md` — the agent's curated long-term memory. - `style-notes.md` — the agent's communication style. Each block starts with TOML frontmatter between `+++` delimiters and requires a `description`. Conditional blocks live in `blocks/conditional/` and are loaded only when [conditions](conditions.md) match. ## Skills Reusable skills live in `skills//SKILL.md`. A skill file has YAML frontmatter with `name` and `description`; the description is shown to the model and the `read-skill` tool loads the full document. ## Memories, tasks, and workspace - `memories/` — files the agent may open, close, and maintain at will. - `tasks/` — scheduled task checklists and related data; `HEARTBEAT.md` drives heartbeats. - `workspace/` — the agent's primary working area and the default working directory for sandboxed commands. The sandbox exposes `/workspace`, `/memories`, `/blocks`, `/skills`, and `/tasks` to command execution.