Create
A named agent with its own rules.
A crew agent is an agent you define once and run again: its own instructions, tools, policy and triggers. Crew is in the navigation rail.
Make one
Choose New agent. You have three starts:
- Start from a role
- A prepared role such as a senior software engineer or a document writer.
- Blank agent
- An empty, least-privilege agent: it may read, list and search files, anything else asks, and anything that changes something is refused until you allow it.
- Describe the agent
- Write what it should do in plain language and choose Generate. You get a draft to review and edit before anything is saved, including its capabilities.
Corenel also ships ready-made agents, among them an orchestrator, a code reviewer and a research analyst. They are listed under Available on the Crew page, and Add puts one in your crew.
What an agent is made of
An agent is a folder of plain files: AGENT.md (its name, description, tags and triggers, then its instructions), SOUL.md (why it exists and what it protects), PERSONA.md (how it talks), POLICY.yaml (what it may do) and settings.json (its model settings). Its system prompt is its job title, soul, persona and instructions, in that order.
Its Location is This project (or This folder), stored in that workspace, or On this machine, which makes it available in every workspace.
The agent page
An agent’s page is one form in sections:
- Identity
- Name, job title, description, tags, location, model and temperature.
- Prompt
- The soul, persona and instructions, with a preview of the system prompt they make. Each has a Draft button.
- Tools
- Allow, ask or deny for each capability, or the same policy as Advanced (YAML).
- Skills
- Which installed skills the agent may load.
- Policy
- Unattended mode, Can call, and Orchestrator.
- Triggers
- What starts it.
- Deployment
- Which paired machines have it, and in which folder.
- Limits
- Max USD, Max tokens and Max time for a run.
Draft the soul, persona and instructions
Draft on any of the three opens a dialog: say what it should say and choose Generate. The draft is written from the agent’s name, role, tools and triggers, and streams into a preview. Draft all three writes every document in order. Nothing changes until you choose Replace or Append; Regenerate tries again.
Rules
What it may do, and who it may call.
Tools and policy
The Tools grid sets each capability to allow, ask or deny, and is saved to POLICY.yaml. A tool from an MCP server is treated as one that changes things.
Unattended mode
When a step would ask and nobody is watching a scheduled run, Unattended mode decides: deny refuses the action (the default), park waits up to 30 minutes for someone to approve, and allow approves it automatically.
Calling other agents
Can call lists the other crew agents this one may hand work to; Allow calling every agent includes agents created later. A delegated call is one level deep: the agent it calls cannot call another.
Skills
A skill is a written procedure an agent can load when it needs it, such as test-driven development or writing a plan. Tick the skills an agent may use in its Skills section; the agent sees a one-line index of them and loads a skill’s full text only when it chooses to. Skills apply when the agent runs on a paired daemon.
Run
By hand, on a trigger, or on a machine.
Triggers
These fire in the browser while Corenel is open:
- Manual
- You click Run.
- File change
- A watched file changes (daemon-side too, when the sidecar runs).
- Run complete
- A run finishes (crew runs are watched daemon-side).
Also Interval (while open), App start and Workspace open.
These need a sidecar started with --daemon:
- Schedule
- A cron time, in UTC.
- Continuous
- Acts on each heartbeat; it needs a budget.
- Self-directed
- The agent schedules its own next wake.
External webhook is listed but not supported yet.
Running and watching
Run fires an agent from the Crew list; Watch follows it. A saved agent’s page has three views: Agent (the form), Chat, and Runs, its history. The Enabled switch pauses its triggers, and Memory shows what it has saved for itself.
Deploy to a machine
When a sidecar is paired, your agents are sent to it, and re-sent when you change them; one started with --daemon runs them. The Deployment section lists every machine this browser has paired with and says, for each, whether its copy is up to date, which folder the agent works in there, and offers Sync now. Only the connected machine’s folder can be changed. Deployment can also run the agent inside a named session.
Find
Tags, search, and asking for the right agent.
Tags and search
Tags are free-form labels (lowercase letters, digits and dashes). Search agents filters the Crew page by name, job title, description and tags, and clicking a tag filters by it. In a chat, ask the agent to find a crew agent for a job: it suggests matches from your crew and the shipped agents you have not added yet, reading only their summaries.