The Agent Roster
One list of every agent that speaks or works for your company, each with a named owner, a last-changed date, and the test it must pass before a change ships.
Jason Lemkin runs SaaStr, a community and events business for software companies. On November 20 he wrote that SaaStr is on track for about $20 million this year "with 3 humans and roughly 20 AI agents in production."
Twenty agents and three people. Every one of those agents speaks for the company to somebody. So who owns each one?
Most companies can't say. Sales set up one agent. Support set up another. Marketing claimed a profile on an outside assistant, and the field team keeps a folder of scripts that sets up new customers. Nobody holds the whole list.
The agent roster is that list. It's the CPO seat's artifact in The Product OS, the way the claim file is product marketing's.
The short version
- The agent roster is one list of every agent that speaks or works for your company.
- Each row has an owner by name and the date the agent last changed. It also has the test the agent must pass before a change goes live.
- The list covers the agent that sells, the agent that supports, and the agent that sets up new customers.
- It also covers the pages a buyer's agent reads, like the docs and the price page. The rule is simple. If an agent does it or reads it, the product team ships it. If a person decides it, sales or marketing owns it. The chief product officer keeps the list.
- Only the product team can change the product and its agents in one release.
What goes on the list
Start with the agents that talk to customers.
The selling agent answers inbound questions and sends the first emails. The support agent answers at 2am. Then there's the company's agent on outside assistants. OpenAI's Businesses in ChatGPT beta lets a company publish a Business Agent there, and the claim file chapter covers what it should say.
The fourth one works for the customer after the sale. It's the deployment agent, the software that sets up a new customer. It connects the customer's systems, loads the starting agents, and runs the first tests. The big vendors are building it now. IBM and Microsoft announced a platform on November 10 that ships with "a fleet of more than 6,000 AI agents and assets." Amit Zavery is ServiceNow's president, chief operating officer, and chief product officer. On the company's October 28 earnings call he said it is building "AI implementation agents."
Last, the pages. Lemkin again: "An agent evaluating your product reads your API docs, pricing page and rate limits, and it never books a demo." No salesperson is in that conversation. The docs, the price page, and the usage limits each get a row, because a buyer's agent treats them as your pitch.
Four columns
Each row needs four things.
The owner, by name. A team isn't an owner. Lemkin passes on a rule from Kyle Norton, who runs revenue at Owner, a software company for restaurants. In Lemkin's words: "Each agent has a person who owns it, and each number has a person who owns it."
The date it last changed. An agent that hasn't changed since the last pricing update is quoting the old price.
The test it must pass before a change goes live. Lemkin describes a model that added a rule nobody asked for and skipped a signed deal. He also caps AI spending at $100 a day, and says an agent went past the cap once. Both are behaviors a test can catch before a customer does. The Eval Is The Spec covers how to write one.
The files it reads. Every customer-facing agent should load the claim file and the current price, and none should carry a pasted copy of either.
Then count the blanks. A row with no owner or no test is where the next surprise comes from.
Who ships what
When agents sell and agents buy, the old line between product, sales, and marketing stops working. Here's the one I use.
If an agent does it or reads it, product ships it. That's every row on the roster.
If a person decides it, sales or marketing owns it. That's the buying committee, the hard objection, the deal that doesn't fit a standard package, the brand, and the room.
The revenue number stays with the chief revenue officer (CRO), the executive who runs sales. The CRO is the selling agent's most important customer and tells product what it has to do. Product builds it, tests it, and ships it. Part 11 of The CPO Transition has the longer argument.
The deployment agent gets a number
The deployment agent's row is the one most companies leave blank. It usually lives as scripts a forward deployed engineer (FDE) wrote for one customer. A forward deployed engineer is the engineer who works on site with the customer.
Give it an owner in product and one number: human hours per go-live. Count the hours people spend between the signed contract and the launch. Then mark each one D or G.
D is a decision. Which work to automate, how the customer's best people make the hard calls, what counts as good enough to launch, and what the product should learn from this customer. Those hours are the engineer's job, and they should stay.
G is a gap. A connection the agent couldn't make, or a test it should have run. Every gap hour is a fix for the deployment agent's owner. If the number stays flat for two quarters, nobody is fixing the agent. Part 17 of The FDE Transition walks through it.
Why the CPO keeps the list
An agent goes stale the same way a help page does. Ship a new plan on Tuesday and the selling agent quotes the old one on Wednesday, unless the two ship together. Only one team can do that. It's the team that ships the release.
The buying side makes this urgent. Cristina Cordova is the chief operating officer of Linear, which makes software for tracking engineering work. In Latent Space's write-up of her interview, agents created 3% of Linear's work a year ago and more than half now. More of your buyers will be software. Software reads every page on that list, and it doesn't call to ask which one is current.
Split the roster across three departments and you get three versions of the product. Keep it in one function and you get one claim file, one set of tests, and one release. That's the practical case for one owner of the loop, and it's the one The CPO Transition keeps making.
What I expect
These are my predictions, not facts.
By the end of 2027, at most AI-native software companies, the selling agent and the support agent ship through the product release process. Each one has a named owner in the product org.
By the end of 2027, at least one large vendor reports the share of launches that had no vendor engineer on site.
Within two years, companies track the share of new customers whose first look at the product was done by an agent. Product owns that number, because product owns the pages the agent read.
One thing to try this week
Make the roster. List every agent that talks to a customer or a buyer in your company's name. Include the chat on the website, the outreach emails, the support bot, anything on an outside assistant, and whatever sets up a new customer.
Next to each one, write the owner's name, the date it last changed, and the test it has to pass.
Count the rows with a blank. Then bring the list to your heads of sales and marketing, and agree on the line together.
Sources: Jason Lemkin, The Top 10 Things That Have Changed in Sales and GTM in the Age of AI, SaaStr, 2026-11-20, SaaStr's figures as he states them and Kyle Norton's rule as he reports it · Basil Chatha with Cristina Cordova, Scaling Linear's GTM to $100M ARR, Latent Space, 2026-11-19, from the written summary · IBM and Microsoft, joint release on Microsoft Source, 2026-11-10 · ServiceNow, Q3 2026 earnings call with Amit Zavery, transcript via Stock Analysis, 2026-10-28
Part of The Product OS, in the CPO seat. The running argument is in The CPO Transition. Related answer: SaaStr sells with 1.5 people and 20 agents. Who owns the sales agent?
Frequently asked
What is an agent roster?+
One list of every agent that speaks or works for your company. Each row has an owner by name, the date the agent last changed, the test it must pass before a change goes live, and the files it reads. The chief product officer keeps the list.
Which agents go on the roster?+
The agent that sells, the agent that supports, the company's agent on outside assistants like ChatGPT, and the deployment agent that sets up new customers. The pages a buyer's agent reads get rows too: the docs, the price page, and the usage limits.
Who owns the selling agent, product or sales?+
Product builds it, tests it, and ships it. The head of sales still owns the revenue number and tells product what the agent has to do. The line is simple: if an agent does it or reads it, product ships it. If a person decides it, sales or marketing owns it.
What number goes with the deployment agent?+
Human hours per go-live. Count the hours people spend between the signed contract and the launch, and mark each one as a decision or a gap. Decisions are the forward deployed engineer's job. Every gap is a fix the deployment agent needs.
Why does the CPO keep the roster?+
Because an agent goes stale the day the product changes, and only the product team can change the product and the agents in one release. Split the agents across sales, support, and marketing and each one tells its own version of the product.
Related reading
Deeper essays and other handbook chapters on the same thread.
SaaStr Sells With 1.5 People and 20 Agents. The Selling Agent Is a Product, and the CPO Ships It.
Agents now sell, support, and buy. What an agent does or reads ships like product under the CPO. What a person decides stays with the CRO and the CMO.
IBM and Microsoft Ship 6,000 Prebuilt Agents and Assets. The FDE's Job Is the Deciding.
IBM and Microsoft ship 6,000 prebuilt agents and assets with engineers attached. When agents do the build, the FDE owns four decisions. Product owns the agent.
Anthropic's S-1 Has No Gross Margin Line. Your AI Product Needs One.
Anthropic's S-1 books compute as opex and shows no gross margin. Where inference sits, and who keeps a 40% price cut, are now CPO decisions.
Anthropic Runs 17 Engineers per PM. The Ratio Is the Wrong Question.
Anthropic runs about 17 engineers per PM and junior tech hires fell from 27% to 9%. The CPO's org question is spans, layers, and who builds the bench.
Your CEO Is Graded on AI. Hours Saved Won't Save Either of You.
Wharton finds product teams lag on AI ROI and most firms count hours saved. The AI story a CPO owes the CEO, and what the CEO should hold the CPO to.
Somebody Has to Steer: The CPO Now Owns the Operating Model
Atlassian's AI builders sped up, then stalled: everyone rowing, no one steering. When prototypes cost a day, the CPO owns the operating model.
Field notes on this chapter