How managed agents work

A Coryntas managed agent is designed around a business role and an agreed outcome. It combines relevant context, tools, and operating controls to carry work across the systems in scope.

A role with a defined outcome#

The role describes the work the agent is responsible for, the information it may use, and the decisions it may take. An invoice exception agent, for example, can have a different scope from an agent preparing a customer renewal review.

The workflow design establishes its permitted operations and when a person must decide. Availability of a tool or connected system does not by itself authorize every action that system supports.

From request to result#

A typical workflow includes these responsibilities:

1. Establish context: identify the request, relevant records, current state, and applicable access.
2. Plan the work: determine the steps and dependencies needed for the agreed outcome.
3. Use permitted tools: retrieve information or carry out the operations within scope.
4. Resolve review points: request a decision when approval or clarification is required.
5. Return the result: provide the deliverable, supporting evidence, and any unresolved work.

The exact execution path, start conditions, and systems are configured for each deployment. Multi-step work can involve specialist agents, but ownership of the overall outcome still needs to be clear.

Assistant and managed agents#

An assistant helps employees ask questions, analyze context, and prepare work. A managed agent has a defined responsibility for a workflow with agreed inputs, action boundaries, and acceptance criteria.

Use the Agent Library to explore examples. Agree on the actual scope with your Coryntas team before relying on a workflow in production.

When work needs a person#

Missing context, conflicting records, sensitive actions, and unexpected system results can require a handoff. The workflow should identify who reviews the issue and what evidence they need.

A request to stop and clarify can be the correct result. It prevents unsupported assumptions from becoming business actions. See the Agent lifecycle for how these cases are evaluated before and after launch.