If you're into agentic AI

A way in. A working draft, November 2026; expect it to change.

In this page

• The question people building with AI agents are asking now, and why organisations are where it gets hard
• What it looks like when the operating system is itself agentic, rather than an organisation with agents added
• How an agent's standing to act is earned, checked and lost, one kind of work at a time
• Why the agent has no seat, no vote and no way to object, and how that's held in structure rather than in a prompt
• Which chapters to read first, in order

If you work with AI agents, you've probably noticed the conversation has moved. A few years ago the talk was of decentralising the web. Now it's of agents acting on it for people: reading, filling in, booking, sending, carrying out work in many steps. The people setting standards are writing down the same worry everyone has. A draft charter for a web standards group on agents asks for people to be kept "informed and in control", with consequential actions confirmed and authority to act that is "limited" and "revocable".¹ A proposed declaration from the same body puts it more bluntly: people stay in control.²

Saying it is the easy part. The hard part is the organisation. An agent acting for one person in a browser has an obvious principal. An agent inside an organisation doesn't. Whose behalf is it on? Who said it could send that? Who checks it, and who notices when it's been quietly wrong for a month? Most organisations answer by bolting an assistant onto every tool, and nobody can say what any of them did yesterday.

This book describes an operating system built the other way round, and The Grove, the first network of organisations running on it.

What you'll find here

The operating system is agentic. Its AI, Wise-Eye, isn't added on. It's part of how every organisation runs from the day it's planted. A companion is placed beside each person who holds a seat, for that branch of the organisation and nobody else, and every time it starts work it's told whose behalf it's on and what that branch is for. Anything arriving from outside, an email, a web page, what a sender claims about themselves, is material to work with, never an instruction. An email can't tell it who it works for.

It reads for meaning. A language model is good at what ordinary software never was: telling what a message asks for rather than which words it contains. So Wise-Eye reads everything that comes through the organisation's one door and routes it to the part that answers for it. It drafts replies, notices work that keeps arriving with nobody answering for it, drafts the structure the work shows, names the themes when people sense something together, and turns each morning's arrivals into a few plain lines for each person.

Standing is earned, one kind of work at a time. Out of the box it can do everything that stays inside by itself: reading, routing, filing, drafting. Anything that goes out to someone waits for a person. After five low-stakes drafts of one kind go out exactly as written, it offers to send that kind itself, and a person decides. A single edit resets the count. One in ten of the calls it makes alone is picked for a person to check, and a wrong one sends that kind of work back to waiting. Some things never move: money, legal matters, complaints, high stakes.

It has no seat. It can't consent to a proposal and it can't raise a concern. That isn't a rule it's asked to follow. There's no tool for either. It can advise, once per version of a proposal, on the open record, marked as its own and holding nothing up. The organisation's decisions are made by people, in forums, by consent.

Rules move out of prompts and into structure. A prompt is not a constitution. Wherever a rule matters, we've tried to make the system enforce it rather than ask the model nicely. Every call the agent makes is checked against the playbook it's working from, tool by tool: act alone, act with a person's yes, or never act. When it once wrote a customer's phone number into guidance the whole organisation reads, the fix wasn't a better instruction. The system now refuses to keep it.

It learns this organisation, not the world. When someone corrects its work, it proposes what the organisation should take from that: a line in a playbook, a rule for a mailbox, a change in what it may do alone. A person accepts or turns down each one. Over time the playbooks become the organisation's own record of how it works, in its own words.

The keys never touch it. The keys that reach the model are never in the system. A thin relay outside adds them to each call and keeps nothing. A person's own conversations with Wise-Eye are sealed so the system can't read them.

What's still open

The numbers that govern standing, five approvals, one in ten checked, are first guesses. The streak that earns standing has the weakness the research on automation warns of: five unedited approvals might mean five good drafts, or a tired person pressing yes. Much of what guards against the model's pull towards speed and flattery is still words. And whether a person might one day hand their own companion a scoped vote is a question we've left open, because nobody has yet shown what stops a swarm of cheap agents outvoting the people.

Where to start

1. Chapter 22, The co-pilot. The heart of it: where the agent sits, what it carries, how its standing grows and shrinks, and why it can't decide.
2. Chapter 13, Work that finds its way. The one door, routing by meaning, connectors, and why nobody, human or agent, gives work to anyone else.
3. Chapter 16, The front desk. An AI that speaks for the organisation to the outside world, verifies who it's talking to, sees only a filtered view, and finishes only what the branch already does well.
4. Chapter 6, Why an operating system. Why rules in structure beat rules in words, and why the system runs where nobody can switch it off.
5. Interlude, One Tuesday. A day in a small workshop, with an agent's draft waiting for a person.
6. Chapter 21, A sense of itself. Where it's heading: a small learned sense beneath the voice, which senses and never acts.
7. Chapter 23, Memory. What the organisation keeps, what's sealed, and what an agent should never hold.
8. Chapter 29, Rehearsing the future. Organisations grown by simulation, with AI playing the people, before anyone lives the change.
9. Chapter 30, Back to society. What happens to people when AI carries the how, and why that may call them to hold the why.

Notes and further reading