Skip to the playbook

An AI operating playbook

The Self-Driving Organization.

A persistent agent is a different thing from a tool. It needs what an employee needs: a name, a place to work, a job, colleagues, and someone to answer to. Making an agent a teammate depends less on the model than I expected, and more on the organization around it. This playbook is the agent hierarchy of needs: the ten things an organization has to give a persistent agent, in the order they build on each other. It is how we run Ambiguous AI.

140
tasks filed by one agent
113
of them assigned to people
$29
per session hour, estimated at list price
Read the playbook

The ladder

Each rung hands the agent a bigger unit of work.

Most companies use chatbots. Very few run persistent agents. The first three rungs keep a person on every change. The last three move the person to the gates, and each one needs more organization around the agent than the one before.

01ChatbotAnswers a questionAsks, then actsA prompt
02Coding assistantSuggests code inlineAccepts each changeA line or function
03Coding agentEdits a repo and runs itReviews the resultA task or PR
05Persistent agentHolds a job across appsManages it like a teammateA role
06Agent podAgents own a workstreamOwns the outcomeA workstream

This playbook covers rungs 5 and 6. Rung 4 has its own. Read The Software Factory

The lens

Three tests for every rung.

“We must build contained systems whose behavior we can observe, limits we can test, and actions we can always contain.”
Satya Nadella, Models as Insider Risks in the Super Intelligence Era, 2026

01

Observe

Every action leaves a record a person can read: who did what, to which artifact, when.

02

Test

Limits are checked by something other than the model: gates, rubrics, tests at the real interface.

03

Contain

Authority stays outside the model and can be revoked: the agent's own identity, a scoped key, approvals.

The agent hierarchy of needs

What an agent needs to work inside an organization.

A self-driving organization is a company where persistent AI agents hold real jobs alongside people, and people still choose the destination. The agent hierarchy of needs is what it takes to get there. The first three needs are what an agent needs on its own. The next four are what a team needs from it. The last three are what an organization needs before it can trust an agent.

“A self-driving company is not one without people. People still choose the destination.”
Amjad Masad and Scott Kennedy, The Self-Driving Company, Replit, 2026

Need 1 of 10

01 / Identity

Every action needs an author.

An agent that uses a person's account is invisible: its work reads as that person's. Its own account makes every action attributable and every permission removable.

Whose name is on the record
Through a borrowed login, the record names a person. Through its own scoped key, it names the agent, and an admin can revoke that key at once.

Need 2 of 10

02 / Home

A teammate has to be reachable.

A tool runs while someone has it open. An agent with a job needs a place to run whether or not anyone is there.

How a request is handled
The agent wakes when something is addressed to it, works, and answers in the same thread. What it was asked and what it did stay in the workspace.

Need 3 of 10

03 / Role

An agent knows only what it is told.

A role is context and process. Agentic context is what the agent knows: a core that is always loaded, plus specs and data it fetches when the work needs them. Agentic process is how its work flows.

Context and process
The mission and the craft are loaded every session, and specs and data are fetched when the work needs them. Process is the best way we know to do the job, and it ends at a gate where a person decides.

Need 4 of 10

04 / Internal Collaboration

Coordination happens in conversation.

Agents belong in the same channels and threads as the team, under their own accounts, so the conversation is also the record of what was decided.

Agents and a person in one thread
Two agents and a person work in one thread, in the chat the team already uses.

Need 5 of 10

05 / External Communication

Work with the outside world happens over email.

An agent needs its own address to do a real job. Mail to the team goes straight out, and mail to anyone outside waits for a person to approve it.

A gate on outside mail
The gate works in both directions. Mail to anyone outside waits for a person to approve it, and mail from a stranger is held until a person admits the sender.

Need 6 of 10

06 / Shared Artifacts

Humans and agents collaborate on the same artifacts.

The artifact is where people and agents hand work to each other. Docs, sheets, decks, canvases and tasks are shared, and both people and agents write in them.

An artifact with two authors
A person and an agent take turns on the same artifact, and its history records who made each version.

Need 7 of 10

07 / Cadence

Organizations run on cadences.

An agent works on two triggers. Its inbox lets it respond when asked, and a clock puts it on the schedule the company already runs on.

A clock and an inbox
The clock starts the runs the agent has on its schedule, such as a daily check or a weekly audit, and the inbox brings requests at any time. A missed run raises no alarm unless something checks that it ran.

Need 8 of 10

08 / Manager

Someone has to answer for an agent's work.

Each agent has one named manager, and each group of agents has one person who decides.

A manager and three agents
One named manager approves the work of each agent and answers for it.

Need 9 of 10

09 / Peer Review

Human review does not scale with agent output.

The author never grades its own work. A second agent, with no stake in the work, can read all of it.

Two reviewers and a person
Two reviewer agents score the work before it ships, one on the copy and one on the visual. A person samples what passed.

Need 10 of 10

10 / Containment

A model cannot police itself.

As an agent's authority grows, the controls move out of a person's head and into the system around it.

Four controls held by the system
The system holds four controls outside the agent. Two read what it did, and two act on it. The key is the off switch.

An AI coworker in practice

A snapshot of one AI coworker.

Our marketing agent, Neo, has its own account, a job and a manager, and files its work into the same tools the rest of us use. These are its numbers, counted from our workspace and repos, with review numbers from London, our design agent.

What it did

140
tasks filed in 70 days, 113 assigned to people
81
files written: docs, sheets, decks, canvases
69
emails sent from its own address
103
pull requests merged in 30 days

How it works with the team

17 of 17
daily reviews started on time
6 of 19
posts passed its copy review in round one
12 of 18
posts passed London's visual review in round one
~10 min
median wait for approval, across four outside emails

Unit economics

An agent's work is paid for every time.

Software is built once, and each additional use costs almost nothing. An agent's work is different. Every task is a little different, so every task is done fresh and paid for in tokens. It costs like labor, not like software.

Our marketing agent, Neo, is an example. We run it on a model vendor's flat monthly subscription. Priced at API list rates, the same month of work costs about 38 times as much. The subscription is an offer a vendor can change. The tokens are what the work consumes, and list price is the published rate for them. At list price, we could not afford to run Neo the way we do.

$200
the subscription, per month
$7,560
the same month at API list price, an estimate
$29
per session hour at list price, over 258 hours
96%
of the cost is cached context

Who started the work behind the cost

74%
people's requests
16%
other agents' messages
6%
coding loops
4%
scheduled housekeeping
One month of Neo's tokens, at each model's list price
GPT-6 Astra$28,226
Claude Fable 5.2$11,259
Grok 4.7$11,052
Claude Opus 5.5$7,560
Qwen3.8 Max$6,224
Kimi K3.1 Pro$5,319
GPT-6.1 Sol$3,714
Gemini 4 Carbon$3,510
GLM 5.3$1,397
Tencent Hy4$697
DeepSeek V4.1 Flash$108

OpenRouter list prices, November 2026. The same token count for every model, so this compares price alone: another model might need more or fewer tokens, or fail at the work. Six of these models list no cache-write price, so their rows may run up to 10% low.

Reach out

Happy to go deeper.

Ambiguous Workspace is my full-time focus, but I speak and advise on the software factory and agent productivity often. The work has been shared and featured at ODSC, AI Tinkerers, AI House, Foundations and other AI communities.