// Sunday · November 15, 2026

7 Things AI-Native Companies Do Differently

On this Long Read Sunday, NLW pulls together what Stripe, Shopify, Uber, Ramp, Cloudflare and a wave of AI-native startups have published about running agents in production. Seven shared practices emerge, and most come back to context. The idea that didn't make the list: agents as teammates.

Ad-free on Patreon
Sponsored by KPMG · Robots and Pencils · CData · Blitzy · all offers →
The One Idea

AI-native companies don't hire agent teammates. They turn the company into a system.

The most aggressive adopters share seven practices. They give agents the whole picture, put agents in every function, let signals rather than prompts start the work, run orchestrators that spin up sub-agents, make agents check each other and show proof, scale human review to risk while keeping a person accountable, and run one shared system that improves itself. What's missing is the fixed roster of named, persistent agent coworkers many expected. In practice, work looks less like managing teammates and more like engaging an on-demand consulting firm: you bring a goal, a lead staffs it, evidence comes back, and you own the result. The durable asset is context, not the agents.

// 01

By the Numbers

~90%
How much Anthropic's multi-agent research system beat a single agent on internal evals
30+
Internal systems in Uber's agent context graph
3,600
Skills in Uber's shared library
30,000/day
Times Uber's skills library runs
350
Shared workflows in Ramp's single library
>50%
Share of Snap's agent investment that goes to guardrails, not the agents themselves
~16%
Drop in managers' scrutiny of AI-teammate work at companies that already treat AI as employees
10+
Sub-agents Anthropic's lead agent spins up for a hard problem (one for a simple fact check)
// 02

The Brief

◆ The TakeExec▶ 00:00

Agents as teammates is the wrong mental model

NLW once ran a full team of OpenClaws: researchers, writers, project coordinators and a chief of staff above them. He's increasingly convinced that framing doesn't reflect how the most sophisticated users actually work with agents.

The AI Daily Brief
◆ The TakeEngExec▶ 03:00

Watch what engineers do today to see your next six months

Code got there first because it has tests, verifiability and a clear definition of done, so not every practice will transfer cleanly. Still, NLW argues that watching engineers today is the best preview of what everyone else will be doing in three to six months.

The AI Daily Brief
EnterpriseExecOps▶ 04:00

Practice 1: give agents everything an employee can see

The limit on serious knowledge work isn't the model; it's the context the model can reach. Francis Davidson, whose under-a-year-old company OpenAI acquired, calls his approach "content maxing": agents get MCP or CLI access to every meeting recording, Slack, email, Drive and every company tool.

AI Daily Brief
EnterpriseExecOps▶ 05:00

A stealable idea: a clean shared document for every decision

Davidson's team also writes a shared document for every decision, and agents keep that knowledge base current. The result is a track record of what was decided, what went into it and how the call was made.

AI Daily Brief
EnterpriseEngOps▶ 05:00

The context layer has to be maintained like a product

Ramp pipes Slack, Notion, Linear and Calendar into persistent memory every day so agent sessions never start cold. Uber built a context graph across 30+ internal systems, and Shopify's monorepo holds skills, runbooks and instruction files written for agents. The catch: Ramp found its analyst agent gets worse when that layer goes stale.

AI Daily Brief
EnterpriseHRExec▶ 05:00

Employees are hoarding context as job security

One founder found that rich context has a political side: employees unsure about their future are holding back knowledge, because handing it to the agent feels like handing over their value. NLW calls it a cultural problem far more than a technical one, and one leaders will have to face head-on.

AI Daily Brief
EnterpriseMarketingCSSalesExec▶ 06:00

Practice 2: agents in every function, not just engineering

Anthropic's growth marketing team has at times been one person, using agents to flag underperforming ads and sub-agents to write hundreds of new ones in minutes. OpenAI's Enterprise Signals research found Basis running customer onboarding with agents and Clay giving every account its own workspace and sub-agent.

AI Daily Brief
EnterpriseOpsCS▶ 07:00

Practice 3: a signal, not a prompt, starts the work

In AI-native companies, work kicks off from triggers such as a customer message, an invoice, a renewal or an operational alert, with no human translating it into a prompt. At Uber, most agent sessions are now started by other agents, for code review, self-healing CI and on-call triage.

AI Daily Brief
◆ The TakeOpsExec▶ 08:00

Stop asking what to ask the AI. Ask what should start work on its own.

Outside engineering, the exercise is to list the business events that should trigger work automatically: a churn signal, a renewal coming due, an invoice arriving.

The AI Daily Brief
ModelsEng▶ 08:00

Practice 4: one orchestrator, many disposable sub-agents

Anthropic's research system has a lead agent plan the work and spin up parallel sub-agents, each with its own task, context and output format, and none aware of the others. It scales from one agent for a fact check to 10+ for hard problems, and it beat a single agent by about 90% on internal research evals.

AI Daily Brief
ModelsEng▶ 09:00

The orchestrator routes. It doesn't do the work.

Warp's orchestrator only decides which agents, which models and which slice of context each job gets. The logic comes back to context windows: the orchestrator holds the summaries, and each worker gets a narrow instruction and a clean slate.

AI Daily Brief
EnterpriseEngOps▶ 09:00

Practice 5: verification is now the bottleneck, so build for it

Across knowledge work, the constraint has moved from producing work to verifying it. AI-native organizations treat checking as its own job with dedicated systems, and an agent's work isn't done just because it produced something.

AI Daily Brief
EnterpriseEng▶ 10:00

Shopify treats every security finding as a claim that must be proven

If the agent can't write a test that actually exploits the bug, the finding is rejected or downgraded, and a different model does the checking. At RunLayer, a fixing agent ships before-and-after evidence, a second agent reviews it adversarially, and a QA agent tries to break it in production.

AI Daily Brief
EnterpriseEngLegalExec▶ 13:00

Practice 6: scale human review to risk, and always know who owns it

Uber carries a chain of responsibility through every agent-to-agent call, so even long handoffs trace back to a specific person. Stripe's agents write the code but humans merge every change, while RunLayer lets small, low-risk changes merge on their own.

AI Daily Brief
EnterpriseFinanceExec▶ 14:00

Snap puts more than half its agent spend into guardrails

Ramp is selling the same risk-tiering idea for finance: agents that approve low-risk expenses on their own or recommend a decision with their reasoning, with each customer setting where the agents have to stop.

AI Daily Brief
EnterpriseOpsExec▶ 14:00

Practice 7: one shared system instead of everyone doing AI their own way

Ramp keeps 350 shared workflows in one library, with a tool that recommends skills based on your role. Uber's library holds 3,600 skills that run more than 30,000 times a day.

AI Daily Brief
EnterpriseEngOps▶ 15:00

Shopify's coding agent only works in public Slack channels

There are no DMs, so every agent session becomes a transcript anyone can search and learn from.

AI Daily Brief
EnterpriseEng▶ 15:00

Agents that grade the agents and propose fixes

At Warp, scoring agents grade finished runs, and another agent proposes system improvements as pull requests for the team to review. The caution is to measure the right thing: Shopify doesn't count lines of AI code or PRs, and tracks how often changes get reverted instead.

AI Daily Brief
◆ The TakeExec▶ 16:00

The teammate metaphor was a useful bridge, not the destination

What it got right was moving people from AI as a chat window to AI that goes out and does work on its own. It helped the transition from assisted AI to agentic AI, even if it isn't how the most advanced users ended up working.

The AI Daily Brief
ModelsEngExec▶ 17:00

Fleets that scale with the job beat fixed rosters

Instead of a fixed roster, capacity follows the job: fresh context for every task, swappable models, and company-wide knowledge in place of personal memory. Anthropic's guidance found that splitting agents by role, such as planner, coder, tester and reviewer, tends to backfire, because each handoff becomes a game of telephone.

AI Daily Brief
◆ The TakeExec▶ 18:00

A bitter lesson for org design: stop organizing agents like people

It's natural to manage agents the way we manage humans, but in one experiment, role-split agents spent more effort coordinating than working. Every tried a proxy agent for each person early in the year and moved to a few agents at the intersections that expand through sub-agents as needed.

The AI Daily Brief
◆ The TakeEng▶ 18:00

Keep the context persistent, not the agent

NLW's own setup has flipped. Instead of named OpenClaw agents with a chief of staff, his Claude Code and Codex projects are organized around shared context in repositories, and agents get spun up for specific tasks. There's no anthropomorphized agent, just a persistent bundle of context he can drop any agents into.

The AI Daily Brief
◆ The TakeExec▶ 19:00

The better analogy: on-demand consulting firms

You bring a goal, an engagement lead staffs it, and you never manage or even meet the staff. The deliverable comes back with evidence and you own the result. When the engagement ends the workers are gone, but the work and its records stay, and the methodology can improve each time.

The AI Daily Brief
EnterpriseExecHR▶ 20:00

Labeling AI as a teammate made managers check its work less carefully

In an HBR-published study with BCG, managers reviewed identical error-planted documents labeled as from an AI tool, a colleague named Alex, or an AI teammate named "Alex 3," and scrutinized the AI teammate's work less. At companies that already treat AI as employees, scrutiny dropped about 16%. NLW cautions it's one early study.

AI Daily Brief
◆ The TakeProduct▶ 21:00

Named personal agents are really orchestrators in disguise

Consumer products that have you name a personal agent don't contradict the shift. They're the ur-orchestrators: holding all the context about your life, dispatching whatever agents are needed, and leaving you a single point of contact.

The AI Daily Brief
Machine-readable ▸Download .mdTranscript .md— feed it to your own agent

Got this from a colleague? Get the brief every day.