Skip to content
HOW IT WORKS

How our AI agents work, and where a person signs off

Packs sit on a directed graph. Manager AI walks it. You see reports. You do not see every hop. We can build the same pattern around one of your workflows, set up with your rules, in accounts you own. Want one for your business? See an AI agent for one workflow (from US$9,550). Deciding whether a marketplace agent is enough? See ready-made vs custom.

HOW A TASK MOVESReaderAuditLanguageGap checkAsset briefCriticReporter

How one task moves through the system

A task is routed across the graph with a spending limit per run. A named person approves anything sent, paid or deployed. What carries into your system is the routing, the spending limits per run, and the human approval steps.

Nothing needs you, so you aren't notified.

What an agent may not do

The brief is the mission. Each run spends only what the task needs. Taste is not a gate.

Spend only what the task needs

Each run has a token budget and stops when it's used up. Finished work isn't re-read.

Only the task it was given

Work only the goal, objects, constraints, and success on the brief. Do not add a slice, channel, or refactor because it looks useful.

Judged against the brief, not opinion

A reviewer step checks each output against the brief and your rules. No extra polish nobody asked for.

What each part does

Each step has a set level of independence: it runs on its own, reports in the digest, alerts you past a threshold, or waits for a person. Legend: Runs alone · Reported · Alerts · Needs approval.

Reader

Read the task. Capture the goal, the objects, the constraints, and the missing info. This pack is not a specialist. It does not rewrite the mission.

Reader: Runs alone

Research

Facts and documents come before audit or copy. They feed the brief.

Research: Runs alone

Audit

Page, UI, UX, design, and analytics are lenses on one pack. Rank them by money and by risk.

Audit: Reported

Language

Understanding, communication, spoken lines, and marketing copy. The critic checks whether claims stay honest.

Language: Reported

Marketing

A gap check, an asset brief, and later generators. Words stay in Language. Publish needs approval.

Gap check: Reported · Asset brief: Reported · Publish: Needs approval

Code

Analyze, implement, test design, parallel runners, and fix. Only the slice on the brief. Deploy sits outside this pack.

Analyze: Reported · Implement: Reported · Test design: Reported · Fix: Reported · Desktop UI: Reported · Browsers: Reported · Mobile UI: Reported · No UI / API: Reported

Critic

Score the artifact against the brief and policy. Stop on a claim break, a failed test, a blown token budget, or a policy break. Do not invent a taste bar.

Critic: Runs alone

Reporter

Run cards, the exception queue, and the digest. This is how you manage without sitting in the graph.

Reporter: Runs alone

Human gate

Production, spend, public send, and a real customer email. These are the interrupts.

Human gate: Needs approval

Deploy

The agent can ask for a release but can't release on its own. Production access stays with a person.

Deploy: Needs approval

Pulse

After launch: crashes, money events, and a regular note; how often is agreed with you. It stays quiet unless the metric moves.

Pulse: Reported

In your company

The same pattern, around one workflow you choose. Sales, ops, or support. Only if it moves a number. You own the code. Off the shelf tools where they fit. Custom work where your expertise is the IP. How AI tools may be used on your material is assessed during discovery. See security.

Method writing lives on Insights. The buyer page for one named workflow is an AI agent with human sign-off. Each piece ends at a fit conversation. It does not end at a case study.

NEXT

Agents that report against your metric.

The first step is a fit conversation. Paid discovery starts only when both sides agree.