Scope is configuration, not prompt wording.
Each agent runs with an explicit grant — the repositories it can touch, the parts of a Blueprint it can alter, the capabilities it can compose, the actions it must hand to someone else. The control plane evaluates that grant at the moment an action is attempted, so an agent cannot widen it by arguing well in its own output. Requests outside scope do not fail quietly; they become escalations, routed to the role or the person who holds that authority. An escalation is the only route out of a scope, and it ends with a person.
Prompts are not a permission system
Instructions in a system prompt describe intent and enforce nothing. Text arriving from a repository, an issue or a document lands in the same window as those instructions and can contradict them. Permissions that matter are checked outside the model, where nothing the model writes can reach them.
Every action is checked at the boundary
An agent proposes one named action: change a field, add a step to a workflow, open a pull request against a specific repository. The control plane matches that action against the agent's scope and the policy profile in force, then permits it, blocks it or turns it into an escalation. The check runs on the action, never on the explanation attached to it.
The worst case is written down
You can state, per agent, exactly what it could do on its worst day. Widening a scope is a configuration change with an author, a timestamp and an approval behind it. Credentials stay with the platform rather than with the agent, and GitHub keeps its own record of what actually landed.
In this capability
- Scope defined per agent
- Checked when the action runs
- Out-of-scope becomes an escalation
- Audited scope changes
AI Workforce
The rest of AI Workforce
Agent Roles
Portfolio Strategist through Release Manager: every role has a written scope, defined inputs and work it is not allowed to do.
Separation of Duties
An agent never approves its own output. Proposal and review are separate roles, held by separate agents, every time.
Model Routing
Choose which model does which job. Two rules hold: no model reviews its own work, and critical workflows use more than one.
Decisions, Not Transcripts
The interface shows the decision, its reasoning, its impact and the options — not a scroll of agent conversation.
Stop managing prompts. Start managing software strategy.
Connect your repositories. Let Foundariq understand what you have already built. Then tell it what your portfolio should build next.
Read-only to start · Your code stays in your GitHub · [email protected]