How Covey works — sandbox, backlog, guard-rails

Identity & Structure

People and agents live in the same org chart — with departments, supervisors and cost centers. Who works where, who reports to whom.

Every agent and every person has a profile with role, task and assignment — the workforce at a glance.

Human roles with finely scoped rights (admin, agent owner, security, auditor, controlling), authenticated via SSO.

Agents work a prioritized backlog like a team — tasks with states, blocks that ask people questions.

Target systems are self-registering plugins with a manifest — dock new integrations without touching the core.

Agents work in real repositories: branches, merge requests, pipelines — full teammates in the development process.

Mailboxes as a target system: agents read and answer mail, triggered by webhooks or polling.

The runtime (first: Claude Code headless) is a thin adapter behind a stable protocol — swappable without changing the platform.

Each agent works in its own ephemeral Docker sandbox with a persistent home. If it is lost, it is rebuilt from config and home.

Never long-lived secrets in the sandbox. Access is brokered at runtime — short-lived, scoped, revocable.

Policies are enforced centrally and outside the runtime, fail-closed. Critical actions wait for human approval.

The sandbox's network egress is controlled — where an agent may talk is decided by the platform, not the prompt.

Every session is recorded: what the agent did, step by step — traceable, with alerts and a kill switch.

Every action carries its cost. Budgets per agent and department, transparent for controlling.

Agents keep a memory of linked Markdown pages with a pgvector index — distilled, not piled up, with scheduled consolidation.

People can share their own knowledge (via the Companion app) with their agents — enforced centrally, fail-closed. In development.