Agents in sandboxes
DELEGATE runs AI agents in sandboxes. Each sandbox is an isolated container with a persistent workspace. The agent works there with a coding engine: Claude Code, Codex or OpenCode.
Agents
Section titled “Agents”An agent is a named definition: its engine, instructions, model and effort, skills, extra commands and the integrations it may use. Agents have versions. You edit a draft and activate it when it is ready; workspaces from the agent use the active version from their next start.
Each agent is shared on its own. The person who creates it is its owner. An owner shares it with people, groups or the whole organization, each with a role:
| Role | May |
|---|---|
| Viewer | see the agent |
| User | also start workspaces from it |
| Editor | also edit and activate versions |
| Owner | also share and delete it |
People who manage agents for the organization see and edit every agent.
Workspaces
Section titled “Workspaces”A workspace is where an agent works. It keeps its files when it stops. You start it, talk to the agent in a chat, and stop it again. A running workspace uses the credentials of the person who started it, and the integrations arrive in the sandbox as MCP servers. The Integrations button of a workspace shows which ones are ready for you and which you still need to connect.
Integrations and credentials
Section titled “Integrations and credentials”Agents reach other systems through integrations: MCP servers and REST APIs (also imported from an OpenAPI document). A credential belongs either to the organization (one account for everyone) or to each user (everyone connects their own account under My connections).
The agent never sees a credential. Every request of a sandbox goes through DELEGATE’s egress proxy. The proxy:
- refuses every destination that is not allowed for the agent,
- adds the credential to the request on its way out,
- checks each call against the rules of the integration, and
- records every call in the audit log.
By default an agent may read. Calls that change something need an approved change set.
Approvals
Section titled “Approvals”The Inbox lists what waits for a decision, for example knowledge revisions. A person who wrote or submitted something cannot approve it (four-eyes rule).
Knowledge and skills
Section titled “Knowledge and skills”Knowledge holds reviewed facts and procedures that agents can look up. Skills are reusable instructions with the tools they need. An agent uses the skills assigned to it.