Skip to main content

Agents

An agent is essentially your own custom LLM with tools attached. Given a request, it intelligently chooses which tools to invoke and in what order — chatting with a user in your custom voice, or running unattended and deciding its next step from each tool's output.

What it's for​

Agents are the workers of the platform. In a chatbot setting, an agent asks the user guiding questions to gather what it needs and invokes tools on their behalf. In an offline setting — a scheduled job or an AutessaDB trigger — it chains tool calls autonomously.

Example: a blog-writing agent on a weekly schedule could scour the internet for trending topics, check your existing blog for what's already published, research its selected topic, and write a draft post — deciding each step from the previous one's output.

From the dashboard sidebar, open Agentic Workflows → Agents. The list has the usual ownership tabs (Owned By Me / Shared With Me / Company Wide) plus search and filters by type, model, and tags. Add New Agent → Create New Agent starts a new one.

Key concepts​

Agent types​

You pick a type at creation (it's fixed afterwards, except when cloning):

TypeBehavior
SimpleYou manually select and configure the agent's tools — the default choice.
MCPPowered by an MCP server; automatically uses all of that server's tools, kept in sync.
CriticIterative refinement: a critic model evaluates each response against your good/bad criteria and sends it back for another attempt until it passes or the max iterations are reached.
SearchPerforms semantic search over an AutessaDB knowledge base table and retrieves matching content.

Type-specific configuration: MCP agents pick their server; Critic agents get a Critic Model, Max Critic Iterations (1–10), and lists of Good Criteria and Bad Criteria; Search agents pick the Knowledge Base Table, the Search Column, Top K Results (1–100), and a Score Threshold (0–1).

Name, description, and instructions​

Name the agent for what it does at a high level — a blog-writing agent might be BlogWriter — and give it a detailed description of when it's useful. The description isn't just for humans: managers read it to decide when to delegate to this agent.

Instructions are the agent's system prompt — the protocol it should always follow. For an autonomous agent, spell out the sequence of steps; for conversational agents this can be lighter since the user steers, but high-level rules still help. Define Prompt Variables (e.g. customer_name) and reference them — along with environment variables — in the instructions as @@{customer_name}; they're filled in at runtime when the agent is invoked. Built-ins like @@{currentDate} and @@{currentTime} (UTC) work here too with no setup — starting instructions with Today is @@{currentDate}. keeps a date-aware agent from guessing the date; see built-in variables.

Model and voice​

The Model Type determines the AI powering the agent: Small Instruction Following, Big Instruction Following, and Big Instruction Following Reasoning are always available, with additional privately-hosted models available to companies with access. Larger models are more capable but use more credits. For spoken interactions, choose a Voice Model and Voice Name — including your own custom voices.

Linked tools​

An agent is nothing without its unique capabilities. In the Associated Tools section, either create a new tool in place or import existing ones. Remember: the tool name, tool description, and argument descriptions are how the model chooses the appropriate functions to call and when — make them as clear as possible. (MCP agents skip this section; their tools sync from the server automatically.)

How to create an agent​

  1. On the Agents page, click Add New Agent → Create New Agent.
  2. Pick the Agent Type, then enter a name (no spaces) and a description of when this agent is useful.
  3. In the editor, set the Model Type, write the Instructions, and define any Prompt Variables.
  4. Attach tools under Associated Tools (or configure the type-specific section for MCP, Critic, or Search agents).
  5. Save — and iterate. The Agentic Copilot can also draft agents for you from a plain-language description.

How to test your agent​

  1. Click Test in the top right of the agent editor. A streaming chat window opens with a live connection to your draft.
  2. Send a prompt or starter sentence and watch it respond — tool calls appear as expandable cards so you can inspect exactly what was invoked with which arguments (and, for Critic agents, each rejected attempt with the critic's reasoning).
  3. If the agent defines prompt variables, a form above the input lets you fill them for the session. You can also attach images and files or record audio input from the microphone.

Testing with audio​

The test window's output toggle switches between Text and Audio responses. In Audio mode the agent's replies stream back as speech in the configured voice and auto-play, with a transcription history panel alongside — so you can hear exactly how your agent will sound before wiring it to a phone number. There's also a separate Real Time Voice button for a live, full-duplex voice conversation.

How to publish​

Editing happens in the draft; when the agent is ready for customers, use Manage → Publish (requires PUBLISH or ADMIN access). Publishing makes this version the default served through the API, and you can roll back any time — see Versions. Manage → Embed generates a chat-widget snippet for your own site, and Share controls teammate and company access.

Sharing and access levels​

Every agent starts private to its owner. Share it with specific teammates — each at their own access level — or company-wide at a single level for everyone. What each level allows:

LevelWhat it lets a teammate do
READView the agent, its versions, and its sharing details, and test-run it
WRITEEdit the draft and submit human evaluations
PUBLISHPublish and roll back versions
ADMINShare and delete the agent

The owner always has ADMIN implicitly; transferring ownership is owner-only.