Build

Agents

An agent is a named AI actor inside a project. It gives a role and identity to model-powered work so conversations, loop steps, memory, tools, access, and history stay connected over time.

Use an agent when you want to recognize and operate the actor, not merely send one isolated prompt. Scout can remain the same agent while its instructions, model, and tools improve and while it works across the app, Slack, and loops.

Two ways to use an agent

Mobius supports two coequal agent experiences:

ExperienceUse it whenDurable record
Agent messagingA person or application should keep talking to the agent.Session
Agent step in a loopThe agent should make one bounded contribution to a repeatable process.Run and session

Messaging keeps the conversation open-ended. A loop gives the work explicit inputs, steps, waits, and a result. The same agent can do both.

Design an agent around one recognizable role

Start with a job a teammate could explain in one sentence: triage support requests, review pull requests, prepare customer briefs, or help users inside a product.

Then make a few deliberate choices:

ChoiceWhat it shapes
DefinitionHow the agent behaves and which model it uses
Toolkits and skillsWhat it can do and which reusable guidance it can load
MemoryWhat selected knowledge follows it across conversations
TablesWhich structured project data it can share
Managed environmentWhether it has a filesystem or repository workspace
RolesThe hard boundary on what it may access or change

Keep the first version narrow. A focused role, short definition, and small tool set make the agent easier to trust and improve.

Identity and definition are separate

The agent is the durable actor. Its definition is the behavior used for work. This separation lets you change instructions, models, time limits, skills, or toolkits without creating a new identity or losing the history attached to it.

Create a second agent when the audience, responsibility, access, or memory boundary differs. Update the definition when the same actor simply needs to do its job better.

Tools, skills, and delegation

Actions let an agent read or change connected systems. Toolkits keep that action set small and understandable. Skills give the agent task-specific guidance it can load when needed.

For focused research or advice, an agent can also delegate a bounded subtask to a short-lived child agent. Delegation is useful when a specialist perspective helps the parent make a decision, but it does not create another durable agent or conversation. The parent remains responsible for the final result.

Memory and shared data

Agent memory holds selected facts and preferences that belong to one agent. Use it for knowledge the actor should carry across sessions.

Use tables when the data belongs to the project and should be shared with other agents, people, or runs. Use a session when the context belongs to one conversation.

Work with agents in the app

Open Build > Agents to create an agent. Its page brings configuration, chat, sessions, memory, tools, skills, messaging, and related loops together.

Start in Chat with a concrete task. Inspect the session, adjust one part of the definition, and try again. Once the behavior is useful, connect a messaging channel or use the agent in a manual loop.

The interactive API reference covers exact agent fields and application operations. The CLI command reference covers terminal use.

Next