Source: raw/7_Things_AI-Native_Companies_Do_Differently.md

Creator: Nathaniel Whittemore (NLW), The AI Daily Brief, a “Long Read Sunday / Big Think” episode · URL: https://www.youtube.com/watch?v=zXrty-ZAIUo · Duration: not recorded · Platform: YouTube (fetched 2026-11-17; sponsor reads for KPMG, Robots and Pencils, CData and Blitzy skipped)

NLW read the public write-ups in which Stripe, Shopify, Uber, Ramp, Cloudflare and several startups describe how they run agents in production, and reduced them to seven shared practices. His summary is that “the company becomes a system”, and most of the seven are about managing context. He then argues that treating agents as named teammates is the wrong mental model and offers on-demand consulting firms as a closer one. Every company example below is his relay of a published post; none of the posts was fetched.

Key Takeaways

  • Give agents everything an employee can see. The limit on knowledge work is the context the model can reach, not the model.
  • Put agents in every function, not only engineering.
  • Let a signal start the work, not a prompt. A customer message, an invoice or an alert triggers the agent without a person typing.
  • Talk to one orchestrator. It picks the workers, the models and the slice of context each one gets.
  • Have a different agent check the work and demand proof. The bottleneck has moved from producing work to verifying it.
  • Scale human review to risk and keep a named person accountable.
  • Run one shared system with a skills library everyone can improve, and measure outcomes such as reverts, not volume.
  • Drop the fixed roster. Capacity follows the job; what persists is company context, not agent identity.

The seven practices, with the examples NLW cites

#PracticeExamples as relayed
1Give agents the whole pictureOdessia (the startup OpenAI acquired; founder Francis Davidson’s post): agents get MCP or CLI access to every meeting recording, Slack, email, Drive and tool, and the company writes a clean shared document for every decision, which agents keep current. Ramp: a daily pipeline pulls Slack, Notion, Linear and Calendar into persistent memory so sessions never start cold. Uber: a context graph built from more than 30 internal systems. Shopify: agents work inside a monorepo that holds code plus skills, conventions, runbooks and instruction files written for agents.
2Agents in every functionOdessia ran agents across strategy, partnerships, product design, marketing, support, legal and finance. Anthropic’s growth marketing was at one point a single person whose agents read CSVs of hundreds of running ads, flag underperformers and write hundreds of new ads in minutes. From OpenAI’s enterprise research: Basis runs customer onboarding with agents, Clay runs account management with a workspace and subagent per account, Exa runs developer-ecosystem growth.
3A signal starts the workRunlayer: work arrives from customer calls, custom Slack channels, failing builds, Slack mentions and pins dropped in the live product. Uber: most agent sessions are now started by other agents, for code review, self-healing CI and on-call triage.
4One orchestrator, workers on demandAnthropic’s research system: a lead agent plans, then starts subagents in parallel, each with its own task, context and output format; one agent for a fact check, ten or more for a hard problem; about 90% better than a single agent on Anthropic’s internal research evaluations. Warp: an orchestrator that only routes, choosing agents, models and context per job. Ramp: a shared coding agent called Inspect.
5Agents check each other and show proofShopify’s security system treats every finding as a claim: if it cannot write a test that exploits the bug, the finding is rejected or downgraded, and the checking model differs from the finding model. Runlayer: an agent opens a pull request with before-and-after screenshots or a recorded walkthrough, a second agent reviews it adversarially, and a QA agent tries to break it in production.
6Review scaled to risk, a person accountableUber: every agent-to-agent call carries a record of the whole chain back to a responsible person. Stripe: agents write the code, a human reviews every merge. Runlayer: small low-risk changes merge on their own. Ramp sells finance agents that approve low-risk expenses and let each customer set the limit. Snap says more than half of its agent investment goes into guardrails.
7One shared, improvable systemRamp: 350 shared workflows in one library, with a tool that recommends skills by role. Uber: 3,600 skills run over 30,000 times a day. Shopify: its coding agent works only in public Slack channels, never DMs, so every session is a searchable transcript. Warp: scoring agents grade finished runs and another agent proposes system improvements as pull requests. Shopify does not count lines of AI code or pull requests; it tracks how often changes are reverted.

Two cautions he attaches to practice 1:

  • The context layer needs upkeep. Ramp found its analyst agent gets worse when its context goes stale.
  • People hoard context. Hebbia’s George Sivulka (captioned “George Savoca”) found employees unsure of their future keep what they know to themselves as job security. NLW calls it a cultural problem leaders must address directly.

Why “agents as teammates” did not hold

  • What the model got right: it moved people from AI as a chat window to AI that does work on its own.
  • What changed in practice:
Teammate modelWhat the write-ups show
Fixed roster of agentsFleets sized to the job
Persistent identityFresh context for every task
One enduring roleModels and workers change
Personal memoryCompany-wide context
  • Splitting agents by job title backfires. NLW cites Anthropic’s multi-agent guidance: planner, coder, tester and reviewer roles turn each handoff into “a game of telephone”, and in one experiment role-split agents spent more effort coordinating than working.
  • Every’s experience: it gave each person a proxy agent early in 2026, then moved to a few shared agents that expand through subagents.
  • NLW’s own setup: he replaced a team of named OpenClaw agents under a chief-of-staff agent with projects organised around repositories of shared context, into which Claude Code or Codex sessions are dropped. “There’s not a persistent named anthropomorphized agent. There’s a persistent bundle of context.”
  • His replacement analogy, on-demand consulting firms: you bring a goal (a prompt or signal); the engagement lead staffs it (the orchestrator); you never manage the staff; the deliverable comes back with evidence; you sign off and own the result; the team leaves but the records stay and the method improves.
  • A cost of the teammate label: a Harvard Business Review study from May, run with BCG, had managers review identical documents with planted errors, labelled as from an AI tool, a human colleague named Alex or an AI team member named Alex 3. Managers checked the AI teammate’s work less carefully; at companies that said they already treat AI as employees, scrutiny fell about 16%. NLW notes it is one early study.
  • Personal agents are not a counter-example. Meta’s Muse and ChatGPT’s Dots act as the single orchestrator a person talks to, which fits practice 4.

Try It

  1. List your triggers. Write down the events in your business that should start work on their own: a churn signal, a renewal coming due, an invoice arriving, a failed build. Wire one to an agent this week.
  2. Write a decision document for every decision and let an agent keep the set current. It gives later agents the reasoning, not only the outcome.
  3. Split maker and checker. Have a second agent, on a different model where you can, verify output and require evidence: a passing test, a screenshot, a source.
  4. Decide in advance which actions need a person, and record who owns each agent’s results.
  5. Move agent work into shared channels or a shared repository so sessions are searchable and skills are reusable.
  6. Change one metric. Track how often agent work is reverted or redone instead of how much is produced.

Open Questions

  • None of the company write-ups was fetched. Figures such as Uber’s 3,600 skills and 30,000 daily runs, Ramp’s 350 workflows, Snap’s guardrail share and Anthropic’s 90% are NLW’s relay.
  • Caption fixes applied: “Odyssey” and “Desi” to Odessia (the wiki’s earlier TechCrunch check gives that spelling); “Run Layer” to Runlayer; “Orp” to Warp ^[inferred]; “George Savoca” to George Sivulka ^[inferred]; “content maxing” is probably Davidson’s “context maxing”
  • The examples are mostly from software engineering, where tests define done. NLW says so himself and expects other functions to follow in three to six months; that timing is his opinion.
  • The HBR study’s sample and design were not checked.