Agents

An agent is a named AI specialist in your workspace, such as an estimator, a procurement clerk or a project coordinator. You give the agent a name, a photo, written instructions (its role, tone and rules), a set of skills (packaged know-how it can load when it needs it) and triggers, short commands like /estimate or #Procurer that call it up. Office admins and operations leads set agents up once. After that, anyone in the workspace can work with them in Chat, in Claude or ChatGPT through the Cooper MCP server, in background AI runs and on calls.

The Agents page lists your workspace's own agents and the System agents that Cooper provides to every workspace. On this page you add, edit, switch on and off, and delete agents.

Key conceptsCopy Anchor Link

Where your agents workCopy Anchor Link

An agent is defined once on this page and then shows up in several places:

To connect Claude or ChatGPT to Cooper, see Connect AI agents.

Open AgentsCopy Anchor Link

1. In the sidebar, open Brain.
2. Click Agents.

The list opens at /connect/agents. You need the Agents permission in the Brain section of your role to see the page. Without it, Cooper shows an access-denied page. If your company's plan doesn't include agents, Cooper shows a page saying the feature isn't in your plan instead. See Permissions.

Read the Agents listCopy Anchor Link

The list shows your workspace's agents and every system agent, with the most recently changed agents first. More agents load as you scroll. The footer shows how many are loaded, for example "Showing 25 Out of 32".

Trigger chips are colored by type: @ mentions in blue, # events in green and / commands in violet.

To hide or show columns, click the Toggle Columns button (the columns icon next to Add Agent) and tick or untick the columns. The list in that menu uses short column names, such as agentMode for Mode. At least three columns stay visible. Cooper remembers your choice for the browser session.

To open an agent, click its row. Your workspace's agents open in the Update Agent panel. System agents, and every agent when you only have read access, open in the read-only View Agent panel.

Add an agentCopy Anchor Link

If you close the panel with unsaved changes, Cooper asks "You have unsaved changes that will be lost if you close this panel. Are you sure?". Click Discard to close without saving, or Cancel to go back to the form. An orange asterisk next to the panel title means there are unsaved changes.

Add triggersCopy Anchor Link

A trigger is a short word that calls up the agent. There are three kinds:

To add triggers, in the Triggers box:

1. Type the trigger.
2. Press Enter, , (comma), Space or Tab. Clicking out of the box also adds what you typed.
3. Repeat for more triggers. A count next to Triggers shows how many the agent has.

Rules:

• If you type a plain word without /, @ or #, Cooper adds @ in front. estimator becomes @estimator.
• A trigger is one word. Pressing Space ends it.
• Typing the same trigger twice on one agent does nothing.
• To remove a trigger, click the x on its chip. Pressing Backspace in an empty box removes the last trigger.
• Each trigger can belong to only one of your workspace's agents. If another agent already uses it, saving fails with "Trigger "…" is already used by agent "…"".

Triggers are not case-sensitive. In background runs, Cooper picks the first agent whose trigger appears anywhere in the request, so pick triggers that are not the start of other words. For example, @tony also matches a request containing @tonya.

Write the instructionsCopy Anchor Link

The instructions are the agent's standing orders. The agent reads them every time it is called up and follows them exactly. Write them as if you were briefing a new colleague: who the agent is, what it is responsible for, how it should talk, what it must never do, and what a good answer looks like.

The Instructions editor opens in Preview mode, which shows the formatted text or "Nothing to preview yet…". To write:

1. Click Edit in the top right of the editor.
2. Type in Markdown. The toolbar inserts Heading 1, Heading 2, Heading 3, Bold (Ctrl+B), Italic (Ctrl+I), Strikethrough, Inline Code, Code Block, Bullet List, Numbered List, Blockquote and Link. Tab indents.
3. Click Preview to check the formatting.

The status bar under the editor counts lines and characters. Instructions have no length limit in the form, but shorter, well-organized instructions are easier for the agent to follow.

Give an agent skillsCopy Anchor Link

Skills are reusable packages of know-how. You build them once on the Skills page and attach them to any number of agents. An agent doesn't read all its skills up front. It sees each skill's name and description and loads a skill when the task calls for it.

The picker loads up to 200 skills. A skill that has been deleted or switched off is left out of the agent's skill list, even if it is still attached.

Choose how and where the agent runsCopy Anchor Link

The Orchestration Settings section controls where the agent can be used and how it runs. Every setting can be left at its default, which lets Cooper decide.

ModeCopy Anchor Link

Preferred modelCopy Anchor Link

The AI model the agent works best with. Inherit from run (the default) leaves the choice to whoever starts the work.

In a background run, a model chosen for the run wins. If the run has none, the agent's preferred model is used, and if that is not set either, Sonnet 4.6. In Chat, the agent uses its preferred model unless someone sets a different one with /model in that conversation.

Thinking effortCopy Anchor Link

How hard the agent thinks before it answers, for example in background runs: Inherit (adaptive) (the default), Low, Medium, High, Xhigh or Max. Higher effort can give better answers on hard problems but takes longer and costs more. Some models don't support every level, and Cooper lowers the level to the nearest one the model supports.

Cost budget (USD)Copy Anchor Link

The most a single background run of this agent may spend, in US dollars. Leave it empty ("Inherit from run") to use the run's own budget, or Cooper's default of $25. A budget set when the run is started wins over the agent's budget. A run that reaches its budget stops.

VoiceCopy Anchor Link

How the agent sounds when it answers out loud in calls and meetings. Pick a voice from the list and click the speaker button next to it to hear a sample. Workspace default uses the first voice on your workspace's ElevenLabs account; it has no sample, so the speaker button is greyed out.

The voices come from your workspace's own ElevenLabs account. If the field says "No voices available. Add an ElevenLabs key in Connected AI Services and they appear here.", an admin needs to add an ElevenLabs key in Connected AI Services first.

Each agent having its own voice helps people tell agents apart in a meeting. A meeting can still use a different voice for itself without changing the agent.

Routing descriptionCopy Anchor Link

"Describe when to route requests to this agent…". When a background run's request doesn't contain a trigger, Cooper's router reads the routing description of every eligible agent and picks the best one. In Claude and ChatGPT, the routing description is also shown when someone lists the agents with @agents.

Write it as a rule, for example: "Use for questions about purchase orders, vendor quotes and material deliveries. Not for invoices or payments." Keep it under 1,000 characters. The router reads only the first 300 characters, so put the most important words first. If it is empty, the router uses the agent's description, or the start of its instructions.

How Cooper picks an agent for a background runCopy Anchor Link

When a background run starts without naming an agent, Cooper chooses one in this order:

1. It looks only at agents that are switched on, not deleted, and whose Mode is Both or Orchestration only. That includes your agents and the system agents.
2. If the request contains one of those agents' triggers, that agent runs.
3. Otherwise, Cooper's router reads each agent's name and routing description and picks the best match.
4. If the router can't match any agent, the run fails with an error that lists the available agents.

To make sure a run goes to a specific agent, include its trigger in the request.

Edit an agentCopy Anchor Link

1. Click the agent's row, or click the ... menu at the start of the row and choose Edit. The Update Agent panel opens.
2. Change any field. The panel has the same sections as Add Agent.
3. Click Submit. Cooper shows "Agent updated successfully." and briefly highlights the row.

Turn an agent on or offCopy Anchor Link

Click the switch next to the agent's name on the list. Green means on. Cooper shows "Agent activated" or "Agent deactivated".

An agent that is off stays on the list with all its settings, but it:

• does not appear in the chat agent picker,
• can't be called up in Claude or ChatGPT,
• is not picked for background runs,
• can't be asked for on calls.

Turn an agent off while you rework its instructions, or to retire it without losing its setup. Hover the switch to see why it is locked: "System agent — cannot be toggled" or "No write permission".

View a system agentCopy Anchor Link

System agents are built and maintained by Cooper and are the same in every workspace. They have a System badge. You can use them like your own agents, but you can't change them.

Click a system agent's row to open the read-only View Agent panel. It shows the agent's photo, name and description, then its Triggers, Instructions, Skills and Orchestration settings (Mode, Model and Thinking). Sections without content are hidden.

If you need a system agent to behave differently, add your own agent with your own instructions and triggers, and switch on only the agents you want people to use.

Delete agentsCopy Anchor Link

Cooper shows "Agent(s) deleted successfully." and removes the agents from the list, the chat picker, Claude and ChatGPT, calls and background runs. Their name and triggers become free to use for a new agent. You can't delete system agents: their checkboxes are disabled and they have no row menu.

To stop using an agent for a while instead, turn it off.

Open an agent from a linkCopy Anchor Link

Links elsewhere in Cooper, such as an activity feed entry about an agent, open the Agents page with that agent's details already showing in a side panel. The panel opens only when the agent is among the agents loaded on the list.

Fields referenceCopy Anchor Link

PermissionsCopy Anchor Link

Access is controlled by the Agents row in the Brain section of a role. An admin sets it in Settings → Roles.

Workspace owners can do everything. Nobody in a workspace can change, switch off or delete System agents; only Cooper's own administrators can, and only they can set the Head agent.

Using agents is controlled separately:

• In Chat, people need Chat → AI Agents in Chat in their role, and chat agents must be switched on for the workspace. See Chat.
• In Claude, ChatGPT, calls and background runs, an agent can only use the tools and records that the person working with it is allowed to use.

Tips and best practicesCopy Anchor Link

• Start from a job, not a model. Name agents after a role people already understand ("Rhonda Procurement", "Site Coordinator") and write instructions for that role.
• Use / triggers. They are the clearest and the form recommends them. Keep them short, one word, letters, numbers and underscores only.
• Put reusable know-how in skills, not instructions. If two agents need the same procedure, make it a skill and attach it to both.
• Always fill in the routing description for agents used in background runs. It is the main thing the router reads. Say what the agent is for and what it is not for.
• Leave model, effort and budget on inherit unless you have a reason. Set a Cost Budget (USD) on agents that run large jobs, so one run can't overspend.
• Turn agents off instead of deleting them while you test changes or when an agent is out of season.
• Give each agent its own voice if several agents join the same meetings, so people can tell who is speaking.
• Check your triggers against the system agents. Your trigger can be the same as a system agent's trigger, and then which agent answers isn't predictable. Look at the Triggers column before you pick one.

TroubleshootingCopy Anchor Link

For AI agentsCopy Anchor Link

Agents are records of the Agent model. Reading and changing them is gated by the Brain → Agents permission (CRM.AGENTS: READ, WRITE, DELETE). System agents (isSystemLevel: true) are read-only for everyone except Cooper super admins.

Rules and behavior:

• Only agents with isActive not false and a status other than Archived can be reached. In-app chat and voice also need agentMode interactive or both; background runs need orchestration or both. talk_to_agent does not check the mode.
• talk_to_agent first matches whole trigger tokens (case-insensitive), then falls back to the first word of an agent's name for @word. Tokens are /, @ or # followed by letters, digits or underscores.
• Background-run routing matches a trigger anywhere in the request text before falling back to the router model.
• Model and budget precedence in a background run: the run's model / costBudgetUSD, then the agent's preferredModel / defaultBudgetUSD, then claude-sonnet-4-6 and $25.
• Deleting an agent sets its status to Archived. Archived names and triggers can be reused.
• Do not set isHeadAgent through db_update; the head flag is managed by Cooper administrators only.

Common errors and fixes:

RelatedCopy Anchor Link

• Brain overview: all the AI features that make up Cooper's Brain.
• Skills: build the reusable know-how you attach to agents.
• AI Team: background runs that are routed to your agents.
• Automations: schedules and events that run agents automatically.
• Org Wiki: company knowledge your agents can read.
• Chat: mention agents in conversations and use agent slash commands.
• Phone and Meetings: calls and meetings where agents answer out loud in their own voice.
• Connect AI agents: use Cooper's agents from Claude or ChatGPT through the MCP server.