Agents

An agent is a specialised worker with its own role, tools, and budget. When it runs, it works in a separate context and hands back a short report. It is not a second copy of the chat you are in.

Agents live under Settings → Harness, with the rest of the Studio harness (skills, rules, hooks).

Some agents are made available out of the box to all.

Agent vs skill

Both live in the harness. They solve different problems.

	Skill	Agent
Is	A procedure (recipe, checklist, template)	A specialised worker with its own brief
Where it runs	Inside the same chat, as extra instructions	In a separate run (a subagent), then reports back
Context	Shares the conversation you already have	Starts clean; only the final report returns
Cost in the main thread	The skill body while it is loaded	Almost nothing until the report arrives
Best for	“Follow this method here”	“Go do this bounded job elsewhere, then summarise”

Use a skill when the main agent should stay in the thread and apply a known procedure.

Use an agent when the work is self-contained, noisy, or worth isolating (enumeration, a side research pass, parallel hosts), so the main chat stays readable.

How the main agent uses them

The main agent is the one you talk to in the composer. In Auto (and custom modes built on it), it can call harness agents as subagents on its own: it picks one by reading each agent’s description, gives it a standalone brief, and continues from the report. Answer and Plan do not offer that.

It does that when:

• The job splits into independent pieces that can run in parallel (several targets, several checks)
• The detail would clutter the main thread (long tool output, noisy scans) and only the conclusion matters next
• A specialist brief, tool allow-list, or cheaper model fits the side task better than the main agent’s full setup

You can also start one yourself: type @ and pick an agent, then the task. That agent runs instead of the main agent for that turn.

@recon enumerate 10.10.10.0/24 and report only hosts with 445 open

TIP

Write the description so the main agent knows when to pick this worker. Vague (“recon specialist”) gets overused. Specific (“enumerates services on a host list and reports open ports with versions”) gets used for what it is good at.

Writing one ProTeamEnterprise

The paid plans include additional agents and allow you to write your own.

Settings → Harness → New agent. Studio opens a form

Field	What it does
Where	Global (host) or a project container. Saving to another place moves the agent there
Name	Short id (for example web-recon). Used for @name in the composer
Description	One line: when to delegate to it. The main agent routes by reading descriptions
Model	Optional. Leave blank to inherit the main agent’s model
Token budget	Stops the subagent once it has spent this many tokens
Allowed tools	Restrict what the subagent may run. None checked means all tools
System prompt	Role and instructions for the tasks you delegate. Edit / Preview tabs

Built-in agents stay read-only. Saving a copy under a free name puts your version beside the shipped one; switch either off in the list to pick which runs.

At equal name, the project (container) version wins over global.

How it feels

When the main agent delegates, the thread shows that a child is working. Expand it to watch the child’s activity. When it finishes, the main agent gets a report and continues. Subagents do not ask you questions mid-run; anything that needs your decision comes back through the main agent and the usual approvals.

What a subagent inherits

Always-on rules apply. Your persona does not: the subagent speaks from its own definition so the report stays a report.

NOTE

Put hard scope in an always-apply rule if every delegated worker must respect it. A persona or a glob rule will not follow them automatically.

Import ProTeamEnterprise

Importing agents is a read-only process on the sources. Staging imported items disabled until you review them is the safe default; you can also import them enabled. Settings → Harness → import.