Agents

What it is

An Agent is an LLM worker that reads a request, decides what to do, calls Tools, and hands over to a person when a rule says so or when it is not sure. You build it in the Agent builder: instructions, a model, Tools, Skills and Knowledge Bases. Find it in the left menu under Build → Agents.

There are two kinds, chosen when you create one: Agent chat answers messages on chat channels, Agent voice takes calls by voice.

Key concepts

• Instructions — the Agent's job description: who it is, what it handles, its rules, when to hand over, its tone. Type @ in the editor to mention a Tool and / to insert a block (heading, list, note).
• Model — the LLM the Agent runs on, picked with Model at the top of the builder. Each model shows its context size and its price in credits per 1M tokens (1 credit = $0.01). A model marked "API key not configured" cannot be used. Thinking level is Auto, Low, Medium or High; Auto means Low in conversations and the model's default elsewhere. A voice Agent's model is chosen by FLAE.
• Tools — the fixed actions the Agent may call. See Tools. Each Tool on an Agent has its own approval setting and settings for this Agent.
• Skills — reusable playbooks the Agent opens when a request matches. See Skills.
• Knowledge Bases — documents the Agent searches and quotes. See Knowledge.
• Variables — inputs given when a run starts, written {{name}} in the instructions. Types: Text, Number, Yes/No, Date, Choice. System values need no declaration: {{today}}, {{now}}, {{contact.name}} and the other contact fields, {{channel.kind}}, {{channel.name}}.
• Escalation rules — cases in which the Agent hands the conversation to a person ("The customer asks for a refund over 2,000,000đ"). They appear when Hand off to a person is on in the Tools list.
• People to ask — Workspace members the Agent may send questions, reports and escalations to, each with what they are Responsible for. With nobody listed, questions go to the Workspace default recipient. Tool approvals do not use this list: they go to whoever started the run (the person testing, the person behind an AI client request, or the Agent's creator in customer conversations). Answers and approvals happen in the Agent Inbox.
• Versions — every publish freezes the draft as a numbered version (v1, v2 …). The header shows "Editing draft v3 · Production v2".

How it works

Agent status

Status	Meaning
Draft	Created, never published.
Published	Has a production version.
Paused	A published Agent someone paused. Resume sets it back to Published.
Archived	Read-only. Cannot be edited or tested. Restore brings it back as Paused (if it was ever published) or Draft.

Pausing only changes the status: channels, Automations and Workforces that use the Agent keep running its production version. To stop them, turn off the Automation or remove the Agent from the Workforce.

Draft and publish

1. A new Agent starts with draft v1. You edit the draft and press Save draft.
2. Publish freezes the saved draft as the production version. Everything that uses the Agent (channels, Automations, Workforces) runs the production version.
3. After a publish there is no draft. Press Create draft to edit to start the next version, copied from production.
4. Discard draft deletes the draft and the test sessions on it; production stays live. An Agent that was never published must keep its draft.
5. On the Versions page you can open any version's settings and Roll back to an older published version. It becomes active again; newer versions stay in the history.

Publish is blocked until: the instructions are written, a model whose provider is configured is chosen, every {{variable}} in the instructions is declared, and every Tool on the Agent is still published. The to fix before publishing button lists what is left and jumps to it. Save your changes before publishing.

How the Agent uses its parts in a run

• It always has the current date and time. Hand off to a person lets it pass the conversation to a person; it is always on for voice Agents.
• With Knowledge Bases attached it gets Search Knowledge automatically. In a conversation (a chat channel, a call or the Test tab) it searches only Public Knowledge Bases; Internal ones are used only in Workforce steps without a conversation.
• It sees the description of each chosen Skill and opens the full text when needed. The Tools a Skill mentions come with it.
• A run stops when it reaches a runtime limit: 30 model calls, 20 tool calls or 600 seconds by default (shown under Advanced → Runtime limits).

Approval of Tools

Each Tool on the Agent has When to ask for approval:

• Ask first — the run waits for a person to approve before the Tool runs.
• Auto — the Agent uses it right away.

A Tool that writes data and that FLAE recommends approving starts on Ask first. If such a Tool publishes a new version after you chose Auto, it asks again until you confirm the choice.

Common tasks

Create an Agent

1. Go to Agents and press New agent, then choose Agent chat or Agent voice.
2. Type a name and, optionally, a description. Write the instructions, pick a model, add Tools and Skills.
3. Press Create Agent. It opens in the builder as draft v1.

Add a Tool

1. In the instructions, type @ and pick a published Tool, or press Add tool in the Tools card.
2. Open the Tools page in the builder's left rail. For each Tool set When to ask for approval, When to use this tool (a note the Agent reads with the Tool's description), and Inputs: Agent fills it or a Fixed value.
3. To remove one, use Remove from agent and delete its @ mention from the instructions.

Attach Skills and Knowledge Bases

• Skills: in the Skills card press Add skill. Only published Skills can be chosen ("Publish this skill before selecting it.").
• Knowledge: open the Knowledge page in the rail and press Attach Knowledge Base.

Declare a variable

Press Insert variable → Declare a variable, or Add variable in the Variables card. Give it a name (lowercase letters, digits and _, starting with a letter), a type, a description, whether it is required, and a default. These names are reserved: contact, channel, trigger, now, today, input, message.

Test the Agent

1. Open the Test tab. Your first message starts a test session on the saved draft (or on production when there is no draft), with the variables you fill in.
2. Write as a customer would. Switch View between As the customer sees it and Details to see tool calls and token usage.
3. Tools run for real here: reads hit the connected App, and writes really happen (after approval when the Tool is on Ask first). Questions to people wait in the Agent Inbox.

Each person sees their own test sessions. A voice Agent is tested with Test call in the browser (allow the microphone).

Publish, pause, archive, delete

• Publish is in the builder header, next to Save draft.
• Pause, Resume, Archive and Restore are at the top of the Agent's Versions page (the clock icon next to its name).
• Delete agent is in the "…" menu on the Agents list. It removes the Agent, its versions, test sessions and Tool runs for good; Automations and Workforces using it stop calling it. Archive it instead to keep it.

Limits and rules

• Who can do what: Workspace Admins and Builders (and Organization Owners and Admins) create, edit, test, publish, archive and delete Agents. Operators can also pause and resume. Everyone else can only view. Nothing can be changed in an archived Workspace.
• Instructions: up to 20,000 characters. Escalation rules: up to 20, each up to 500 characters. Variables: up to 20. People to ask: up to 20.
• Tools: at most 20 per Agent, counting the Tools its Skills bring and its built-in Tools. Skills: at most 20. Knowledge Bases: at most 10.
• Auto approval can only be set by a person in the app. The Assistant can build and change draft Agents, but it cannot publish them or turn approval off.
• The Evals tab is a preview: suites you add there are not saved, a run does not test the Agent (every case shows "Not run"), and it never blocks Publish. The Memory setting under Advanced is not available yet.
• Model prices are shown in credits per 1M tokens; 1 credit = $0.01.

Related

• Tools — what an Agent can call.
• Skills — playbooks shared by several Agents.
• Knowledge — documents an Agent searches.
• Workforce — several Agents and Tools on one canvas.
• Agent Inbox — where people answer an Agent's approvals and questions.
• Automations — put a published Agent to work on Messenger.
• Runs — what the Agent did in each conversation.
• Chat with the Assistant — have the Assistant draft an Agent for you.
• Apps — the App actions behind Tools.