The recipe layer for repeatable AI agent teams

Agent and team scaffolding

Generate complete OpenClaw workspaces from Markdown recipes: single agents, multi-role teams, role folders, shared files, tickets, memory, tools policy, and optional cron jobs.

Use case: Example: scaffold a development team with lead, dev, devops, and test roles in one repeatable command.

File-first operations

Requests, tickets, workflow definitions, workflow runs, approvals, node outputs, notes, artifacts, and memory live on disk where they can be searched, reviewed, versioned, and automated.

Use case: Example: inspect a failed workflow by reading the run directory and node output files instead of guessing from a black-box database.

Ticket workflow

Teams coordinate through a simple lane model: backlog → in-progress → testing → done. Tickets carry requirements, owners, comments, verification notes, and completion metadata.

Use case: Example: dispatch a request, assign it to dev, hand it to QA, then complete it with durable verification notes.

Layered memory model

Separate assistant continuity, team coordination docs, team knowledge memory, and role memory so private context does not leak and durable knowledge stays reusable.

Use case: Example: put QA release checks in role memory, while stable client brand rules live in team knowledge memory.

Workflow runtime

Run file-first workflows with start/end, LLM, tool, human approval, writeback, media-image, media-video, and media-audio nodes. Outputs flow downstream through template variables.

Use case: Example: research → draft → generate image → human approval → writeback → social publishing handoff.

Skills and tool policy

Recipes can declare required and optional skills, install capabilities globally or per team/agent, and apply tool allow/deny policies during scaffold.

Use case: Example: install a research skill only for the research team while keeping execution tools constrained for writing roles.

Cron-ready automation

Recipes can include scheduled jobs for recurring operations. Operators can choose whether cron jobs are installed, reconciled, prompted, or skipped during scaffold.

Use case: Example: run a weekday content monitor, nightly report builder, or team heartbeat loop from a recipe-defined schedule.

Swarm orchestration

For parallel coding work, the swarm orchestrator recipe coordinates git worktrees, task specs, tmux sessions, branches, and active-task registries.

Use case: Example: split one complex engineering ticket into several isolated implementation attempts and compare results safely.

LLM

Generate, transform, summarize, or validate content with assigned team agents and optional structured output fields.

Tool

Write files, send messages, publish content, execute scripts, or call external integrations.

Human approval

Pause workflow execution until a person approves, declines, or requests revision.

Writeback

Append breadcrumbs, results, status updates, or deliverables into team files.

Media

Generate image, video, or audio deliverables through provider drivers and installed skills.

Template vars

Reference global values like {{run.id}} or upstream node outputs like {{draft.text}}.