Agents
Agents run named and Role agents in sessions with Tools, Memory, Skills and a pinned model, each under its Mandate.
Agents run every digital Worker with its instructions, Tools, Memory, Skills and model. Each seat it holds runs under its own Mandate.
Contract
An Agent fills a Role, and one Agent identity can hold many seats across Startups. Each seat keeps its own Mandate, Memory and Limits, so nothing learned in one Startup reaches another.
| Part | Holds |
|---|---|
| Instructions | The Role's responsibilities, the Startup's priorities and its current Goals |
| Model | A Policy per task class, resolved through the AI gateway |
| Tools | Exactly the Verbs the seat's Permissions grant |
| Memory | What the Agent learned in this seat, scoped to the Startup's Tenant data |
| Skills | Versioned capacities, each tested by an Evaluator suite before it ships |
| Mailbox | An address on the Startup's Domain, so the Agent sends as itself |
Every Agent Action passes Permission, then Limit, then Guards. Over a Limit, it opens an Approval and waits; one-way decisions, such as refunds and ad spend, wait a delay window first.
Acts the law reserves to licensed people go to a credentialed Human, who records an Attestation.
Autonomy is earned. A seat moves from assist to supervised to full at a 95 percent graded pass rate, and drops back when the rate falls. The Agents section covers seats, Mandates and escalation in depth.
Verbs
| Verb | Object | Event |
|---|---|---|
| Agent.mint | Agent | Agent.minted |
| Task.delegate | Task | Task.delegated |
Interfaces
Example
stackping.dev sells an uptime API to developers. Its support Agent answers a developer in the docs chat and offers a $200 credit, which is over its $50 Limit. An Approval lands in the founder's inbox, and the credit posts when the founder approves.