What an AI-native company has to make explicit

The September 6 episode of The AI Daily Brief treats "AI-native" as an operating model rather than a layer of software added to old work. Host NLW works through Alex Lieberman's 30 features of an AI-native company, a list shaped by Lieberman's work at 10xLabs and his earlier experience founding Morning Brew. 1

The episode's argument has a practical center: a company has to make its work legible to agents, give agents usable context and skills, measure what they produce, and keep a human owner attached to the result. The list is broad, but several features form one chain. Process maps become context. Context supports repeatable skills. Skills feed workflows. Evals and ownership determine whether the workflows deserve more autonomy.

Lieberman's first feature is a function-by-function blueprint of how the business works. NLW gives the reason: important details often live in people's heads, hallway conversations, and side chats rather than in operating manuals. An agent that lacks those details has to reconstruct the work from scratch. A process map gives the agent a starting context for redesign. 1

NLW adds a qualification that keeps the idea from becoming a simple automation program. Agents may discover a better way to reach a goal than the way people currently work. A blueprint should therefore describe the goal, constraints, decisions, and exceptions. The blueprint should leave room for a new workflow instead of forcing an agent to imitate every old step.

The next layer is a daily work harness for every employee. Access to a frontier model is only the beginning. A work harness also organizes context, reusable skills, and tool access. The employee learns how to give an agent the material it needs and how to let the agent act within a defined boundary. 1

The episode places shared context near the center of the company. Lieberman's third feature calls for an intelligence layer that brings together structured and unstructured data, documents, and business logic in a queryable form. NLW prefers a mesh or lattice for a large organization, where multiple sources of truth can connect and agents can inspect differences between them. The underlying need stays the same: agents need organized context before they can perform dependable work. 1

"Treat context as code" gives the idea an operating habit. Architecture documents, conventions, and planning material need maintenance because agents build on those artifacts. The company has to record information worth learning from, attach feedback to a specific output, and keep permissions in the data layer. These practices turn context from background reading into a maintained part of the workflow.

The same context work also affects cost. The episode describes a design in which an agent loads the slice of company knowledge it needs rather than the whole repository. Metadata can tell the agent whether a file is relevant before the agent spends context on its contents. The detail sounds technical because token use becomes a direct operating cost once agents run repeatedly. 1

The episode separates skills from prompts. A prompt can tell an agent what to do once. A shared skill can carry a repeatable way of working across a team and across several steps in a workflow. A skills-distribution system can trigger the same instructions where they belong, manage agent behavior, and improve token efficiency. 1

That distinction leads to a larger change in software delivery. Agents can plan, write, test, review, and ship code while people define intent and acceptance criteria. Planning can use a higher-effort model, while execution can use a faster and cheaper model. Model routing becomes part of the architecture, matching task difficulty with model capability and measuring the cost of a successful result. 1

The same pattern applies to non-engineering work. A workflow becomes self-improving when the company gives an agent a goal, guardrails, and an objective measure of success. The agent can run the process repeatedly, inspect the result, and adjust the next run. A sentence such as "the interface should look good" gives the loop a weak finish line. A measurable test gives the loop something it can evaluate. 1

Lieberman's list includes a "Citizen Developers SDLC": a governed path that lets non-engineers take a small solution from an idea to production while preserving access control, versioning, and software conventions. NLW draws a boundary around the idea. Marketing and HR staff do not need to become replacement software engineers. They gain the ability to build tools for their work and contribute clearer specifications to product and engineering teams. 1

That shift also changes the role of governance. The episode describes legal, HR, and IT teams as transformation partners that help design policies around the new work. Permissions, approval paths, and guardrails should be built into the system so every workflow does not reopen the same debate. Autonomy can then rise through stages: observation, suggestion, action with approval, and action inside a defined boundary. 1

The episode also places people at the beginning and end of many workflows. NLW calls this a human sandwich: a person sets the work in motion and judges the result, while agents handle the middle where the steps are measurable. The right intervention point in the middle depends on the process, so the phrase works as a design question rather than a fixed template. 1

The episode closes with a comment that NLW says was widely shared by listeners: every AI workflow needs a clear owner, a measurable goal, and a person responsible when the workflow fails. That requirement gives the 30 features a management spine. Context, skills, routing, evals, and guardrails improve the machinery. Ownership decides whether the machinery is producing the outcome the company actually needs. 1

A practical reading of the episode follows the same order. Map one workflow. Give the people involved a reliable harness. Gather the context and skills that workflow needs. Set a finish line that an evaluation can check. Add approval and permission boundaries. Then measure the result before expanding the pattern.

Lieberman's list is a set of claims about where companies may go as agents become more capable. The episode's durable point is narrower and more useful: an AI-native company is a company that makes work, context, feedback, permissions, and ownership explicit enough for agents to participate. The technology changes quickly. Those operating responsibilities remain attached to the people who own the result.