Connect an AI service through the supported tools, including MCP. Use your own agent harness or integrate another provider. The workspace remains the governed body of knowledge on which the agent works.
Working with agents
Give an agent a defined job, scoped knowledge, and a reviewable place to work.
Choose the AI service
Give the job a boundary
An agent's context identifies its workspace, use case, persona, permitted actions, and spending authority. Executable Use Cases gives those actions a business definition and testable contract.
For example, an agent checking an application receives the information and actions needed for that assessment, rather than unrestricted access to every company workspace.
Review the result
Use a branch for proposed changes. Inspect what the agent changed and require the applicable checks and approvals before publication. Branching supports the same process for human contributors.
Use hand-over to continue through the people and agents responsible for each step.
Task and journal records help trace the work, its inputs, and outcome. These controls bound authority and expose evidence; they do not make an AI model infallible. The MCP guide explains the supported connection.
Developer reference
Technical details
Server-enforced capability scope
An admitted run binds principal, workspace, branch, persona, Use Case, permitted Stories and budget. Tool calls execute inside that authority envelope; generated prompt text cannot widen it.
Model Context Protocol exposes tool discovery and invocation. MCP transport authentication is not a workspace grant. The receiving service authorizes the requested operation against the run context.
Stories declare typed input/output concepts. Implementations execute those interfaces rather than accepting arbitrary model-generated database privileges. Proposal commits remain isolated until destination validation and required approval authorize publication. Task/transaction IDs correlate tool execution with the resulting diff.