Agents

Agents help create the operating system and use that software to operate the business. The Continual agent is built in. Authorized external agents can use the same tools, business rules, and permissions instead of relying on a separate copy of how the company works.

What you can ask

Start with the result you want, the context that matters, and any decisions that should stay with a person. Continual can help organize the system, run the work, or improve what already exists.

The Continual agent

People work with the agent through conversation in Continual or a connected channel. The same project identity, history, tools, and permissions follow the conversation across channels.

The work queue keeps durable items visible until someone resolves them:

• questions and missing information;
• exceptions the agent should not resolve alone;
• approvals for consequential actions;
• proposed source or behavior changes.

Pending judgment must not disappear into old conversation text.

External agents

An authorized MCP-capable agent can use the operating model's published objects, tools, documents, and other capabilities. Continual may expose narrow project services through the same endpoint, but business logic remains in the backend. Customers can therefore use a specialized external agent without copying their business context and logic into it.

“Any agent” means any compatible client using the capabilities its authenticated principal may discover and invoke. It does not mean anonymous access or exposing every internal platform operation.

Every call carries an authenticated person or delegated-agent identity and a project scope. Permissions, approvals, idempotency, provenance, and audit remain server-side. The gateway handles identity, discovery, routing, and invocation records; it does not own business logic.

Agents that build

A building agent reads the repository and proposes source changes: a new object, a changed rule, a document, a skill, or an app.

What it produces is a change: an intent, a diff, and evidence. You review it the way you would review a colleague's work, except the evidence comes attached.

Because the backend is code, every change can be reviewed, tested, and reverted.

Agents that operate

An operating agent triages incidents, chases missing documents, reconciles invoices, drafts customer replies, and escalates what it should not decide.

It acts through the same tools a person or app calls:

This makes its work:

• Authorized. An agent has an identity with roles and permissions.
• Attributable. Material tool invocations record the actor and provenance.
• Inspectable. Its roles and available tools define what it can do.

Building and operating are modes of work, not separate agent products. The same Continual agent may propose a source change, run an operation, or ask a person for judgment.

Practical guidance

• Give the narrowest role that lets the agent work.
• Gate what is hard to reverse: money leaving, a customer being contacted, or a record being deleted.
• Put durable pending judgment in the work queue rather than leaving it only in conversation.
• Convert important corrections into feedback and evaluations so the same mistake does not recur.