How AI Native Corporations Really Function

Expensive subscribers,

I’ve spent the previous few months interviewing leaders at AI-native corporations. I’m now satisfied that:

Onboarding and managing AI brokers IS the job, it doesn’t matter what your operate is.

On this free deep dive, I wish to share how three AI-native corporations —Linear, Ramp, and Manufacturing unit — put this precept into follow. Listed here are some quotes from every:

1. Nan Yu (Head of Product, Linear): “You’ll have AI workforce members that you may assign duties to and speak to similar to the way you speak to individuals.”

2. Geoff Charles (CPO, Ramp): “In the event you’re not utilizing Claude Code, it doesn’t matter what your function is, you’re most likely underperforming.”

3. Eno Reyes (CTO, Manufacturing unit): “We codified product administration, frontend UI, knowledge evaluation, and extra into reusable abilities that any worker can invoke.”

Learn on for an inside take a look at how every AI-native firm operates.

Linear’s strategy to brokers is formed by their product. Nan (Linear’s Head of Product) believes that:

Brokers needs to be first-class workers. You need to be capable of add them to tasks, assign them to points, and point out in feedback.

Nevertheless, Nan additionally believes that the human at all times stays accountable for the result. Right here’s how Linear builds merchandise with brokers in each step:

1. Perceive the issue. Brokers learn and summarize each buyer dialog from Intercom, Zendesk, and Gong. They auto-create points, de-dupe them towards the backlog, and assign them to the fitting workforce.

2. Establish the answer. Since brokers have entry to buyer conversations, they may help you iterate on a spec by pulling data-backed insights from a number of channels.

3. Make a plan. Brokers can break your spec into concrete tickets and route them to the fitting groups robotically. At Linear, brokers now create the vast majority of tickets. The human’s job is to overview their work and modify context over time.

4. Execute. Bugs and small options get assigned on to brokers like Codex and Cursor. For advanced options, engineers launch Claude Code and use Linear MCP to tug in full difficulty context.

From Nan:

Need to construct with brokers like Linear? Listed here are 4 sensible steps that Nan shared on what your workforce can do right this moment:

1. Each developer ought to default to a number one agentic coding device. That is the best first step. Present the official device (Cursor, Claude Code, or Codex) and handle it so you possibly can see utilization.

2. Complement with an async cloud coding agent. Async background brokers can one-shot most small adjustments and bug fixes. Cursor and Devin have good choices right here.

3. Insist that designers and PMs work instantly on the codebase. Brokers like Claude open a low-friction path for PMs and designers to make adjustments instantly within the codebase. Everybody ought to try to be a builder.

4. PMs and entrepreneurs ought to default to an AI interface. These features needs to be doing 80-100% of their work by way of a chat interface — whether or not that’s Claude, ChatGPT, Notion AI, or one thing comparable.

Nan sees a future the place people will collaborate with brokers on the spec degree — defining what must be constructed and why — after which passing issues off to brokers to deal with every little thing downstream.

If Linear exhibits learn how to make brokers a first-class a part of your workforce, then Ramp exhibits learn how to get your total firm to undertake them.

In 2025, Ramp shipped 500+ options, reached $1B income, and did all of it with 25 PMs.

They did this by requiring each single operate (eng, product, design, gross sales, advertising and marketing, authorized, finance) to onboard and work with brokers.

Geoff (CPO Ramp) shared a framework for evaluating AI proficiency for each worker that I discover extraordinarily sensible:

1. L0: Typically makes use of ChatGPT. These individuals will most definitely not be on the firm long-term. In the event you’re not a self-starter with a development mindset towards AI instruments, Geoff says it’s going to be very onerous to coach you to excel.

2. L1: Makes use of and tweaks GPTs, tasks, and inner AI instruments. These individuals are experimenting with AI however haven’t automated actual work but.

3. L2: Builds an app that automates a part of their job. These individuals can commit code or give significant suggestions on different individuals’s work utilizing AI instruments.

4. L3: Methods builders. These individuals are constructing the AI infrastructure and abilities that speed up everybody else on the workforce.

The corporate’s aim is to push everybody up the ladder. L0s self-select out. L1s turn out to be L2s. L2s turn out to be L3s. And L3s affect the remainder of the group.

Geoff additionally shared 5 steps that any firm can take to turn out to be AI-native:

1. Take away all friction. Give entry to fashionable AI instruments with no constraints on tokens or budgets and create an inner repository of AI abilities anybody can pull from. If the setup is difficult, most individuals gained’t undertake.

2. Make adoption seen. Create public Slack channels the place individuals can share what they constructed. In any respect-hands, showcase non-engineers doing spectacular issues, like finance constructing a treasury system or advertising and marketing automating web site creation.

3. Present hands-on help. Host workplace hours that anybody can be a part of to construct AI abilities and workflows. Have designated AI specialists whose total job is to evangelize, get individuals arrange, and assist them attain the “aha second.”

4. Monitor utilization and intervene. Ramp tracks token consumption throughout AI instruments per worker. Management shares this knowledge to create pure accountability and step in when somebody’s utilization is low.

5. Make it a hiring requirement. PM interviews now embody a devoted session the place it is advisable construct a working product after which clarify why you constructed it and the way it works.

Geoff summed up his management philosophy for each function at Ramp in a single line:

From Geoff: “If I inform my workforce 10 occasions that the CTA must be above the fold, the repair isn’t saying it the eleventh time. It’s encoding that suggestions into an automatic design crit course of or AI talent in order that it by no means occurs once more.”

If Linear and Ramp present how corporations undertake AI brokers, Manufacturing unit exhibits what occurs while you construct round them from day one.

Manufacturing unit is a 55-person AI software program improvement firm valued at $300M that’s structured round AI from the bottom up. Right here’s what makes them totally different:

Manufacturing unit doesn’t rent PMs and engineers individually. As a substitute, they rent product engineers who handle and work with AI brokers. A typical day appears like:

1. Look at traces from agent runs to see the place the system made poor selections.

2. Write fixes not as code, however as governance (e.g., an replace to a talent, a brand new lint rule, or a refined automation)

3. Assessment solely the PRs that brokers flag as high-risk (brokers care for the remainder).

4. Recommend new concepts and debate prioritization with colleagues and AI.

This work isn’t straightforward and requires deeper experience, however the leverage is big.

Brokers want a codebase they will truly work in to be efficient. Manufacturing unit scores codebases throughout 5 maturity ranges, and Degree 3 (“Standardized”) is the place most groups have to intention first.

As soon as your codebase is agent-ready, the subsequent step is giving brokers the information to make good selections by way of abilities (principally simply textual content markdown recordsdata).

Manufacturing unit makes use of abilities to encode knowledgeable and firm information into one thing that any agent or worker can use. Right here’s an inventory of abilities that Manufacturing unit makes use of internally and hyperlinks to their markdown recordsdata so that you can copy and modify:

1. Product administration talent. Product ideas, an 11-star expertise framework (borrowed from Airbnb’s Brian Chesky), PRD template, scoring rubric, and language steerage multi function markdown file.

2. Frontend UI integration talent. Instructs Droid on learn how to construct options utilizing Manufacturing unit’s design system, routing conventions, and testing requirements.

3. AI knowledge analyst talent. Run exploratory evaluation, construct visualizations, and generate statistical stories utilizing the complete Python ecosystem.

4. Inside instruments talent. Construct admin panels, help consoles, and operational dashboards with correct entry controls and audit logging baked in.

5. Vibe coding talent. Quickly prototype new net apps from scratch with fashionable frameworks.

In the event you can encode what your greatest individuals know into abilities, you don’t want to rent specialists for each operate.

To recap:

Listed here are six issues you possibly can put into follow proper now:

From Linear:

1. Default each developer to an agentic coding device like Cursor, Claude Code, or Codex.

2. Get PMs and designers into the codebase. Allow them to submit PRs and ship code utilizing brokers. Cease routing each small change by way of an engineer.

From Ramp:

3. Measure AI proficiency throughout your workforce. Ramp’s 4 ranges framework offers you a shared vocabulary for the place individuals are and the place they should go.

4. Monitor AI utilization and make it a part of efficiency expectations. You may’t enhance what you don’t measure, and incentives matter.

From Manufacturing unit:

5. Rating your codebase for agent readiness. Use Manufacturing unit’s agent readiness ranges to grasp whether or not your codebase is prepared.

6. Codify your workforce’s experience into reusable abilities. Encode what your greatest individuals know into talent recordsdata and make it straightforward for each people and brokers to make use of them.

Above all, onboard brokers such as you’d onboard a human. Give them context, join them to your operational stack, and hold a human accountable for his or her outcomes.

Let me know what you suppose within the feedback!