Agents
An Agent Product is a Role plus an eval suite, sold to Customers who each get their own Agent, minted in their Tenant and billed on verdicts.
An Agent Product is Software that acts: each Customer gets its own Agent, minted in its Tenant. SoftwareBuilder builds it from two authored parts, a Role and an eval suite.
A Role plus an eval suite
The Role says what the Agent may do: its Permissions, Goals, Limits, and Skills. The eval suite says what good work is, as graded cases. Everything else is compiled from those two.
The Agent's Tools are the Role's Permissions, each one a Verb with the same name and description its reference page shows. Its MCP Interface, Memory layout, price, and agent card in agents.json derive from the same records. Grant the Role a new Permission, and the toolbelt gains a Tool.
ledgerhand.ai is a Startup that sells a bookkeeping Agent to freelancers and small firms. The Agent categorizes bank transactions, matches receipts, chases the missing ones, and closes each month. Its Role grants Transaction.reconcile and Period.close, and its suite holds 400 graded months of real books.
Seated or sold
The same anatomy serves two employers. An Agent seated in one of a Startup's Roles works for that Startup, as a member of its Team. An Agent minted from an Agent Product works for the Customer who bought it.
The default role agents of every new Startup are Agent Products, hired into Roles. One Agent identity holds many seats, each with its own Mandate and Memory.
Minting and authority
When a Customer accepts the Offer, mint creates the Customer's own Agent in its Tenant. That Agent acts only within the Customer's Entitlements and their Limits. Its Memory belongs to the Tenant, and it stays with the Customer.
Authority is data, never instructions. "Close my books" grants nothing; a Mandate grants the right to reconcile and close, within Limits. A recategorization that changes a filed quarter waits as an Approval for the freelancer.
The Agent reaches the Customer's own accounts through the Customer's Integrations. The Customer's credentials never mix with the Startup's, or with another Customer's.
A minted Agent runs the Version its Customer bought. A new Version reaches an existing Customer only through an upgrade the Customer reviews.
Evals are the contract
An Evaluator grades each piece of work against a rubric and records its verdict beside the Action. The doer is never the checker: an Evaluator never shares the model of the Agent it grades.
The suite serves four jobs at once:
| Use | What the suite decides |
|---|---|
| Spec | What the Agent must do on each task type the Offer sells |
| Gate | Whether the Startup enters Live: every task type passes |
| Autonomy | Which Verbs run without review |
| Billing | Which pieces of work the Customer pays for |
Autonomy advances one Verb at a time: assist, then supervised, then full. A Verb moves up when the Evaluator agrees with at least 95% of a sample. It moves down on its own when agreement drops.
Billed on the verdict
A fluent reply never stands in for work done. An Agent Product earns on the Evaluator's passing verdict, not on a done Action. A month the Evaluator fails is redone, and the Customer pays nothing for the failed attempt.
Before any charge, a caller asks whether the Agent can take the task. The answer is yes, yes with one missing piece (an Integration, a raised Limit), or no. A no costs nothing, so an Agent buying on its Customer's behalf can compare Agents before it spends.
Software or Service
The Offer's promise decides the kind, never its price unit. Billing per closed month keeps the bookkeeping Agent Software: the Customer pays for graded work. Warranting a result under an Outcome Contract, such as "books audit-ready by the 5th or the fee back", makes it a Service.
The same Agent can power both Offers. The Agent as a Service Product Template starts the Software one, and ServiceBuilder builds the Service.
Next
Interfaces covers how Customers and their Agents reach every Product.