Identity and mission
Assign a stable ID, a responsible team, a role, and a measurable outcome. A narrow mission is easier to evaluate than an open-ended instruction.
PLATFORM / AGENT RUNTIME
An agent should be more than a prompt with tools. Define its mission, owner, permissions, memory, budget, and evaluation criteria. The runtime concept turns these controls into the operating conditions for each task.
02 / THE OPERATING CONTRACT
Assign a stable ID, a responsible team, a role, and a measurable outcome. A narrow mission is easier to evaluate than an open-ended instruction.
List approved systems and actions explicitly. Read, draft, and execute are different privileges and should be reviewed separately.
Scope memory to the worker’s role. Keep source references and freshness alongside retrieved context, rather than treating remembered data as current truth.
Track changes to policies, tools, model configurations, and agent instructions. Compare evaluations before promoting a new version.
03 / CONNECTED BY DESIGN
Inspect the sequence.
Each stage has a defined responsibility.
Select a node to inspect its role in the workflow.
Define the source of truth, permitted actions, accountable owner, and acceptance criteria. Scope matters more than an open-ended instruction.
Check the outcome, capture unresolved exceptions, and preserve the relevant policy and evidence. A task is complete only within its defined boundary.
04 / A CONCRETE EXAMPLE
This fictional worker belongs to Northstar’s finance operations team. It uses a private reasoning model concept, can read ERP and warehouse records, and can draft a report. FIN-04 requires human approval for an external action. Its mission is reconciliation; payment approval is outside its role.
A practical agent specification includes task limits, data permissions, timeout behavior, failure recovery, and a named escalation owner.
05 / QUESTIONS & BOUNDARIES
Architecture descriptions show intended system behavior. Validate implementation, integration, data handling, and operational controls before relying on a production workflow.
The design principle is that permission changes require a separate authorized control process. An agent’s request does not grant itself access.
The proposed runtime pauses or stops the task and records the limit reached. It should not silently continue using more time or resources.
Identity, context, policy, evaluation, verification, execution, and audit form one operating model. Explore the related components below to see their responsibilities.
06 / CONTINUE THE EXPLORATION