Agents: each task passes a second model, or it is redone or free

The reference for every call on this page is Integrations.

Seats your agents fillCopy Anchor Link

Each seat is a hub Role with a written contract: its purpose, results, target, standards, limits and channels. An agent Worker fills it. Put a person in the seat instead and the contract stays the same.

One seat is never an agent's. You, as head of sales, set the number, answer every approval the other seats raise, and take the first sales calls yourself.

The marketing seats sit beside the sales ones. A Marketing lead picks Plays, a Content lead writes for search and AI answers, and Paid media spends only what you authorized: a standing approval (a Mandate, set with Mandate.authorize).

Switch seats on as customers arriveCopy Anchor Link

Every seat can run from day one, but each one you switch on adds what it asks to your queue. Switch on the seats that move the number at your stage, and keep the queue to what you answer in one day.

Four seats join on their own signalCopy Anchor Link

Some seats follow a signal, not a customer count. The AE joins when booked calls pass 5 a week and your next open slot is more than 2 business days out. The SE joins when the first buyer asks for a security questionnaire or a test on their own data.

The Marketing lead joins when a second Play runs beside cold email, so each one is read against its number every week. Paid media joins when a term forecasts under your break-even cost per click, $4.29 at $149 a month, and you authorize the spend (Planning estimate).

An agent is a Role, a Function and its ToolsCopy Anchor Link

An agent Worker fills a Role: its Permissions, Goals and limits. It runs as an agentic Function with a pinned model, instructions, Memory and Skills.

Its Tools are exactly its Role's Permissions. The calls on Data and Headless SaaS are therefore the whole of what it can do, and each one is logged.

Swap the agent for another, or for a person, and the Permissions, the limits and the record stay as they were.

Agent tasks you buyCopy Anchor Link

An agent task is the hub's Agent kind from four kinds: one piece of work that passes its Evaluator's verdict, or is redone or free. It bills on that verdict, never on effort.

The Evaluator runs on another model and grades the output against what passing means for that task. A reply or a winning ad is the market's choice, so these tasks warrant the work, not the result.

Each is checked by a second model; a miss is redone, or it is free.

It acts as you on every Venue you authorizeCopy Anchor Link

Your agent acts as you on every site you authorize, in one of three ways. Where a site has no API for a step, a browser agent signs in as you and does it.

You connect each account once with Integration.connect and Handle.connect. Every Action then records whom it acted for, under which standing approval, and how it ran.

What waits for youCopy Anchor Link

Agents send and post inside their seat's limits. One approval rule on a class, such as send, makes every action in it wait for you, and going over a limit always asks.

You see the plan and its price before anything runs. Granting a standing approval is itself your call: when an agent asks for one, it waits for you. Each class and its default amount: the hub's approvals page.

// 25 cold emails a day from each warmed inbox on tryleaseline.app, for 30 days.
// The Mandate holds 50 a day across both; each Mailbox’s own pace holds it to 25.
await $.Mandate.authorize({ agent: 'agents/sdr', verb: 'Contact.enroll', channels: ['mailboxes/warm-1', 'mailboxes/warm-2'], limit: { count: 50, period: 'day' }, expiresAt: '2026-11-26T00:00:00Z' })

# The SDR seat’s Role file: one approval rule on send
approvals:
  - { class: send }

To see the first 20 before they go, run the enrollment with try, and it returns the drafts. To make every send wait for your OK, set one approval rule on send.

Work is checked, not trustedCopy Anchor Link

Every agentic step has a checker that is not its doer: an Evaluator on another model, a rule in code, or a person. The hub's anatomy of a Service sets that rule, and the weakest check sets the assurance.

Claims get an adversarial review: every fact in a first line must match the row's evidence, or the line is cut. Quality gets an affirmative review, which asks what would make the draft better and applies it before it counts.

Action.evaluate grades any Action against a rubric and records the verdict as an Event, so a Play reads pass rates the way it reads replies.

Bring your own agent over MCPCopy Anchor Link

Your own agent connects to the MCP server with one key and gets three tools. search finds a call by the job you describe, fetch reads its page, and do runs it.

It works under the same Role, standing approvals and approval rules as the agent it replaces. The hub's bring your own agent covers the swap.