The AI Daily Brief: Artificial Intelligence News and Analysis · · Transcript & summary by WhipScribe
Concrete examples from Stripe, Shopify, Uber and Ramp show how AI-native teams run agents in practice, and why the teammate metaphor may mislead you.
Tap a chapter to open that moment in the transcript; the link to it is copied for you.
“AI as teammates was a useful mental model to help people make the transition from assisted AI to agentic AI.” NLW · at 16:53 — open in the transcript →
“Each handoff becomes a game of telephone.” NLW · at 17:37 — open in the transcript →
The host argues that the main limit on useful agent work is context, not the model. Companies that have published their setups, including Stripe, Shopify, Uber and Ramp, give agents broad access to meetings, chat, email and internal tools, and they keep shared records of decisions and their reasoning. That context needs upkeep, because stale information degrades agents, and employees who fear for their jobs may hold back what they know. The same companies put agents into every function, not just engineering, and start work from signals such as an invoice, a renewal or an alert instead of a manual prompt.
Teams are also moving toward one orchestrator agent that breaks a job into narrow sub-tasks and hands each to a worker with a clean slate. Verification is treated as its own job: one model checks another's findings, agents attach screenshots or tests as proof, and human review is scaled to the level of risk, with a named person accountable for each result. Consistency comes from shared skill libraries and public, searchable session transcripts, and success is measured by outcomes such as how often changes are reverted rather than by lines of code.
The host argues that the idea of agents as named teammates with fixed rosters and lasting identities mostly failed to match practice. Anthropic's guidance found that splitting agents by job role created costly handoffs, and a Harvard Business Review study found managers checked work less closely when it was attributed to an AI teammate. He proposes thinking of agents as on-demand consulting firms: bring a goal, let a coordinator staff it, check the deliverable and keep the records. Personal agents from Meta and ChatGPT may fit that model as orchestrators.
NLW says that a handful of aggressive AI adopters, including Stripe, Shopify, Uber, Ramp and Cloudflare, have published how they run agents in production, and that a common set of seven practices is visible across them. He argues that the most useful lesson is to watch what engineers do today to see what other organisations will be doing in three to six months.
“These are companies like Stripe, Shopify, Uber, Ramp, Cloudflare, as well as smaller AI native companies and startups as well.”NLW · 2:04
NLW says Uber built a context graph from more than 30 internal systems so that its agents have the information they need. He argues that for meaningful knowledge work the limit is usually the context, information and systems an agent can access rather than the model itself, and that this context layer must be kept up to date.
“the limit tends not to be the model, and it hasn't been for some time.”NLW · 3:44
NLW says Ramp runs a daily pipeline that pulls Slack, Notion, Linear and Calendar into persistent memory, so agent sessions never have to start cold. He also reports that Ramp's analyst agent becomes worse when its context layer gets out of date or is neglected.
“A daily pipeline pulls Slack, Notion, Linear, and Calendar into persistent memory, so agent sessions never have to start cold.”NLW · 5:05
NLW says Anthropic's research system uses a lead agent that plans the work and spins up sub-agents in parallel, with one agent for a simple fact check and 10 or more for a hard problem. He reports that this approach beat a single agent by about 90% on Anthropic's internal research evaluations.
“Anthropics research system is one of the clearest public examples.”NLW · 8:14
NLW describes a Harvard Business Review study published in May, of researchers working with BCG, in which managers reviewed documents with planted errors labelled either as AI tool, human colleague Alex, or AI team member Alex3. Managers checked the Alex3 work less carefully, and scrutiny dropped about 16% at companies that already treat AI as employees.
“The study looked at researchers working with BCG.”NLW · 20:09
The episode argues that AI-native companies are best understood through seven practices: giving agents full context, putting agents in every function, starting work from signals, using an orchestrator with sub-agents, having agents verify their work, scaling human review to risk, and keeping one shared improving system. The host concludes that the agents-as-teammates model is the wrong one and that on-demand consulting firms are a closer analogy.
The episode lays out seven practices seen at AI-native companies: give agents the full context an employee could see, put agents in every function, start work from signals instead of prompts, use one orchestrator with sub-agents, have agents check and prove their work, scale human review to risk, and run one shared system that improves over time. The host draws these from public write-ups by Stripe, Shopify, Uber, Ramp and Cloudflare.
The host says that for meaningful knowledge work the limit is usually not the model but the context, information and systems the model can access, and that this has been true for some time. This is why companies are giving agents access to meetings, chat, email, drive and internal tools. The host notes that the context layer then needs its own management, because out-of-date information makes agents worse.
The host argues the teammate model assumes a fixed roster of agents with set roles and lasting identities, while companies are moving to dynamically sized fleets that scale with the work, fresh context for each task, and company-wide knowledge rather than personal memory. He also says a Harvard Business Review study found managers scrutinised work less when it was labelled as coming from an AI teammate.
According to the host, Ramp runs a daily pipeline that pulls Slack, Notion, Linear and Calendar into persistent memory, so agent sessions never have to start cold. The host also says Ramp's analyst agent gets noticeably worse when that context layer is out of date or neglected, which is why maintaining it is a continuing job.
The host describes a Harvard Business Review study published in May, of researchers working with BCG, in which managers reviewed the same documents with planted errors. The only change was the label on the author: an AI tool, a human colleague named Alex, or an AI team member named Alex3. Managers checked the Alex3 work less carefully, and at companies that already treat AI as employees, scrutiny dropped about 16%.
The host says Uber's shared skills library holds 3,600 skills and runs more than 30,000 times a day, which makes workflows reusable across the whole organisation. Shopify takes a different route: its coding agent only works in public Slack channels and never through direct messages, so every session is a transcript anyone can search and learn from.
The host says an engagement works like this: you bring a goal, a coordinator staffs it, you never manage the staff, the deliverable comes back with evidence, and you sign off and own the result. When the engagement ends the workers are gone, but the records and methodology stay and can improve. He calls it not perfect but directionally closer than teammates for understanding how these systems work.
Our own transcript of the public episode, with a timestamp about every 90 seconds. Tap a timestamp or the link button on a paragraph to copy a link that opens this page there.
0:00People have been excited about the idea of agents as teammates for a lot of 2026, going all the way back to the OpenClaw explosion. I myself had an entire team of OpenClaws, some researchers, some writers, a bunch of different coordinators for projects, and a chief of staff that sat above them all. Increasingly though, I am getting more and more convinced that agents as teammate is in fact the wrong mental model. Or at least, it doesn't at all reflect how the most sophisticated users are using agents in practice. Today we're going to get into that, but also look at more broadly, seven things that AI-native organizations are really doing, and why so many of them come back to how they manage context. The AI Daily Brief is a daily podcast and video about the
0:41most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Robots and Pencils, CDATA, and Blitzy. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors at ai daily brief.ai. One more last call to register for the AI champions bootcamp. It's the first time that we're running this program. It is six live sessions with new far. It kicks off tomorrow and we'll transform you not only into an agent user, but into someone who can help agents propagate across your entire
1:22organization. You can find info about that at besuper.ai. At this point, we have all been drowning in China's spy and chip export control scandals for way more time than I think anybody wants. And so for this week's Long Read Sunday slash Big Think episode, I really wanted to do something that was once again, much more practical. And as you might imagine, one of the things that I am constantly watching for is companies reporting how they're actually using AI. I think one of the great things that we all struggle with is that there's so little information relative to the speed at which things changes about how these new capabilities are actually finding their way into existing systems. You hear a lot of companies aspire to be AI native,
2:04but what does that actually mean? Well, luckily, over the past few months, You hear a lot of companies aspire to be AI native, but what does that actually mean? Well, luckily, over the past few months, at least a handful of the most aggressive AI adopters have started publishing exactly how they run agents in production. These are companies like Stripe, Shopify, Uber, Ramp, Cloudflare, as well as smaller AI native companies and startups as well. Individual operators, of course, are posting their setups on X on a much more frequent basis. Now you read enough of these things and a picture starts to become clear of what, at least at the moment, the closest to AI native organizations have in common.
2:34Today, we're going to look at seven of those practices. Practices which theoretically you could put in place as soon as you're done listening to this. And then we're going to spend time on one thing that I might imagine many of you to expect to be on the list, but it isn't, which is the concept of agents as teammates. I've been getting increasingly convinced that the agent as teammate mental model is an incorrect one. And I'm starting to see some evidence for other people finding something similar as well. Now, one important caveat up front is that, as you might guess, a lot of the examples that we'll discuss today come from the world of software engineering. That's not only because code got
3:08there first, it's because of why code got there first. It has tests, verifiability, done has a clear definition. And in some cases, applying these things to other areas outside of coding and software engineering won't be as clear. Still, I've always said that one of the best ways to actually see into the future of your own organization or your own AI practices is to go watch what the engineers are doing today to give you a sense of what you might be doing in three to six months. So let's get into these seven practices and then talk about why agent organization seems increasingly less like a team of coworkers and more like hiring consulting firms on demand. So let's start with these seven practices for getting useful work out of agents.
3:44And if you had to give it a TLDR, the idea is that the company becomes a system. The first practice, which will come as a surprise to exactly zero of you, is about context. Context which has lurked around as the important unspoken issue for basically the entire history of AI. As we all know so well, especially for significant, meaningful knowledge work, the limit tends not to be the model, and it hasn't been for some time. It's the context and information and systems that the model has access to. So the first practice of the most advanced AI-native organizations is to give agents the whole picture, to give them everything an employee can see, or in some cases, even more than a single employee, because agents are working so cross-functionally. One example, which by the
4:28way was one of the inspirations for this piece as a whole, came from the blog post by Francis Davidson after OpenAI acquired his company Odesia. This is a company that was less than a year old, and people were really excited to see Francis spell out a ton of his different practices. Now, when it comes to this idea of giving agents everything an employee can see, Davidson called the approach content maxing. And what it means is that agents get MCP or CLI access to everything. Every meeting recording, Slack, email, drive, and every tool the company uses. Now, one thing that I thought was really notable, it is another stealable idea that many of us might not be doing, is that Odessia also wrote
5:05clean shared documents for every decision. Agents keep that knowledge base up to date, but it provides a track record of not just the decisions, but what went into them and how they were decided. Ramp has something similar. A daily pipeline pulls Slack, Notion, Linear, and Calendar into persistent memory, so agent sessions never have to start cold. Uber built a context graph from more than 30 internal systems. And at Shopify, agents work inside a mono repo effectively. They have one repository that holds the code plus skills, conventions, runbooks, and instruction files that are written for agents. Now, one of the key catches in giving agents this amount of rich context, holding aside, of course, any security concerns, which I am blissfully breezing over
5:43right here, is that this context layer is in and of itself something that has to be managed. Ramp's analyst agent, for example, they found gets absolutely worse when its context layer gets out of date or is neglected. One additional fascinating challenge that George Savolka of Hebbia found was that this actually has some political issues to it as well. Basically, employees who are unsure about their future are starting to hoard context as job security. If you know what your value is, why would you hand it to the agent? Now that is, of course, a cultural problem much more than a technical one, and one that I think leaders are going to have to deal with head on. Practice number two, put agents in
6:19every function, not just engineering. Being AI native does not mean having an engineering team who is AI native. It means being AI native across the organization. Odessia, again, put agents across strategy, partnerships, product design, marketing, support, legal, and finance, and the different groups share back in the way that they use agents to understand how that might differ from other groups as well. Over in Anthropic, their growth marketing team is, or at least has been in the past, a single person. They've built a robust set of workflows that leverage agents to do things like reading CSVs of hundreds of ads that are running, flagging the underperformers, sub-agents for writing headlines and descriptions, and the net result is that they can write hundreds of new ads in minutes.
6:57OpenAI's enterprise signals research found examples like Basis running customer onboarding with agents, Clay running account management with a workspace and sub-agent for every account, and Exa using agents to run developer ecosystem growth. I recently did my episode about how I use agentic video pipelines, which is another example of agents outside engineering. With practice number three, the description gets a little more abstract, but I think what it is will become clear. The idea is that a signal, not a prompt, starts to work. In other words, in non-AI native companies, AI isn't doing something unless someone explicitly tells them to do it. In the more AI native companies, the work starts without anyone telling it to do it
7:35because it's based on some trigger that when it happens doesn't require a human transfer that trigger to a prompt. The types of signals or triggers include things like a customer message, an invoice, a renewal, some sort of operational alert. At RunLayer, they find that work comes in from customer calls, custom Slack channels, builds failing, Slack mentions, and pins people drop right in the live product. In Uber, most of their agent sessions are now started by other agents around things like code review, self-healing CI, and on-call triage. And if you want to think about this outside of engineering, just try to think about which events in your business should start work on their own. A churn signal, a renewal coming due, an invoice arriving.
8:14The question changes here from what should I ask the AI to do to which things happen that should start work automatically. Practice number four starts to get into some of what we'll talk about in the team member section. But the point of this one is that increasingly, teams are not interacting with a whole bunch of standing agents. There is one main or orchestrator agent that can figure out what's needed and can spin up or call in other sub-agents as whatever challenge it's dealing with requires. Anthropics research system is one of the clearest public examples. A lead agent plans the work and then spins up sub-agents in parallel. Each one gets its own task, its own context, and its own output format, and doesn't even know the others exist. The number
8:54scales to the job. One agent for a simple fact check, 10 or more for a hard problem. This approach to agents beat a single agent by about 90% on Anthropics internal research evaluations. ORP has an orchestrator whose only job is to route work but not do the work itself. For each job, it picks which agent or agents, which models, and which slice of context the agents get. RAMP has a coding agent that they call inspect that's a good example of how this is coming into the shared or team agent domain. And one of the key reasons for this again comes back to context windows. Rather than having to maintain comprehensive context across a large number of different agents, the orchestrator keeps the summaries and the key context, and each worker gets
9:32a narrow instruction and a clean slate, and theoretically just what they need. Practice number five, having agents check each other's work and show proof. At this point, I think it's pretty clear that the bottleneck across pretty much all types of knowledge work has moved from producing the work to verifying it and or just making sure it doesn't suck. AI native organizations are not only viewing verification as its own important job, but are also designing specific systems for that step and that layer. They don't view the agent's job as being done when it produces whatever it was supposed to produce. So a couple examples that have been published. Shopify has a security system that treats every
10:09finding as a claim that has to be proven. If it can't write a test that actually exploits the bug, the finding is rejected or downgraded. And the model that checks the finding is different from the model that found it. At run layer, an agent fixes any problems and opens a pull request with before and after screenshots or a recorded walkthrough. A second agent then reviews that adversarially. And then even after it ships, a QA agent tries to break it in production. So to sum up, we've got one, check the work, and two, Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding,
10:59evaluating, and refining its outputs. These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at kpmg.com slash US slash AI amplifiers. The best teams don't have a single star carrying everyone else. They know their own strengths and each other's weaknesses and play to both. That's the team Robots and Pencils has built on purpose. Nobody there is grinding through busy work to pad a headcount number. People come for the hard problems and they stay because everyone around them is leveling up at the same time. In a market full of companies that are just trying to hire fast, that's worth a look. Check out robotsandpencils.com slash careers.
11:42This episode is brought to you by CDATA. Token spend is climbing, security is holding agents back, and companies rolling out AI initiatives might not know how it's truly being used. The root cause? AI is cut off from the systems and context your business runs on. That's why CDATA built Connect AI Gateway, one control point between your AI and your systems. Connect AI Gateway is not just another static model router. Prompts get intelligently routed to the right model with exactly the tools and data it needs governed down to the record. The result? Accurate answers at a fraction of the token cost with a context engine that learns your business with every interaction. Built on the connectivity layer powering Palantir, Google,
12:20Salesforce, and hundreds more. Give it a try and start your free trial at cdata.ai. That's C as in connect, data.ai. Every AI coding tool on the market does the same thing first. It starts writing code. Blitzy does the opposite. Before writing a single line, Blitzy spends days reverse engineering your entire code base. Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade. The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building. Other tools guess at context with grep searches and markdown files. Blitzy never guesses. It builds true understanding
12:57first, then delivers over 80% of entire software epics autonomously. Validated, end-to-end tested, production-grade pull requests. That's why Fortune 500 engineering teams trust Blitzy with the codebases that matter most. See for yourself at Blitzy.com. That's B-L-I-T-Z-Y dot com. Practice number six, scaling human review to the level of risk involved and keeping someone, a person, accountable. The point is that AI-native companies neither review everything nor do they review nothing. They try to decide ahead of time what types of things are going to need a human, and they always know which human owns the result. Whenever there's an agent-to-agent handoff at Uber, they retain a chain of responsibility that traces back to a specific person.
13:44Every time one agent calls another, the call carries a chain of responsibility that traces back to a specific person. Every time one agent calls another, the call carries a record of the whole chain, so even a long agent-to-agent handoff traces back to that responsible person. At Stripe, while agents are writing the code, it's humans who merge it. Every merge has a human reviewer. At RunLayer, they do it a little bit differently. Small, low-risk changes merge on their own, and humans step in when asked. Ramp is actually selling something similar for finance, offering finance agents that approve low-risk expenses on their own or recommend a decision with their reasoning. Each customer gets to decide how much those agents can do alone and where they have to stop.
14:16Interestingly, Snap says they put more than half of their agent investment into guardrails rather than the agents themselves. The seventh practice is about the consistency of the system across the organization. Most companies, even who are working very hard to use AI well, still tend to have wild variations in how different people use AI. Basically, everyone is using AI in their own way, or maybe at least different departments are using AI in their own way. The more AI native the organization, the more there tends to be a single system that's dynamic enough to adapt and allow everyone to improve it while still working within the same system. And there are a couple hallmarks of this. Ramp and Uber, for example, both have shared skills
14:54libraries that make workflows reusable across the whole organization. Ramp has 350 shared workflows in one library with a tool that recommends skills based on your role. Uber's library, meanwhile, has 3,600 skills that runs over 30,000 times a day. Another way to build this sort of shared system comes from Shopify. Its coding agent only works in public Slack channels, never through DMs, which means that every session is a transcript that anyone can search and learn from. These AI-native companies are also figuring out novel, interesting ways to help the systems improve themselves. For example, at Warp, scoring agents grade finished runs, and another agent proposes improvements to the system as pull requests the team reviews.
15:33Basically, they're using some of their agentic power to examine the agentic system in an ongoing way. One caveat or caution within this is to make sure to measure the right thing. Shopify, for example, does not count lines of AI code or pull requests. It focuses instead on things like how often changes get reverted. So these are seven practices that tend to be hallmarks of AI-native organizations. And again, as I said at the beginning, the one that isn't in here, at least not in the way that we were talking about it a few months ago, was agent teammates. Now, I don't want to be overly pedantic or semantic about this, but I do think that exploring why the way we work with agents has evolved a little bit differently than many of us thought is useful, even if it risks
16:17either A, overstating the case, or B, trying to ascribe a single reality when lots of different individuals and lots of different companies are going to have different experiences. I do think it's fair to say that many folks thought that agents as teammates would be the defining feature of AI-native companies, that every person would have their digital representation, or that there would be specific agents that have specific roles that are maintained and persistent, and that you anthropomorphize through names and things like that. Now, what this did get right is that it helped us shift our thinking from AI as a chat window, that we could ask for help for work that we do, to something that could go out and do work on its
16:53own. AI as teammates was a useful mental model to help people make the transition from assisted AI to agentic AI. But there are a few different reasons why it didn't exactly play out that way. One is that a teammate model tends to assume some fixed roster. When we're just using the analogy of agent as digital version of the people that work for you, that means a set number of agents, each with a role they own, and an identity that lasts. Instead, what we've seen is a move towards dynamically sized fleets of agents that can scale up or scale down based on the work at hand. So what that means is, instead of a fixed roster, you have capacity that follows the job. Instead of a persistent identity, you get fresh context for every task. Instead of a
17:37single enduring role, models and workers can change. And instead of personal memory, the most important thing is actually company-wide knowledge, company context. In fact, interestingly, Anthropic's guidance on multi-agent systems has found that splitting agents by job function, like planner, coder, tester, reviewer, tends to backfire. Each handoff becomes a game of telephone. In one experiment, agents split by role spent more effort coordinating with each other than actually doing the work. I think a little bit of this is a version of the bitter lesson. It's very natural for humans to assume that the right way to organize agents is to approximate the things that we've learned about managing people, when in fact what's better is to just let them figure out how to organize themselves,
18:17because those structures might look very different than what works best for us. I think that these games of telephone and effort coordinating are the chief example of that. Another example of this shift came from Every, who did a bunch of experimentation towards the beginning of the year with each person having their own proxy agent, and then realized that didn't make sense at all. It made much more sense to have a smaller number of agents that lived at the intersection and could expand themselves through sub-agents as was required. And every week that goes on, you see more examples of this shift popping up as well. I've certainly experienced it. At the beginning of the year when I was messing around with OpenClaw, I had a research agent, actually multiple different research agents based on
18:52different projects. I also had a coordinator agent. I also had agents who were responsible for keeping track of different projects and then a chief of staff agent that sat on top of them. Now my behavior is pretty much not like that at all. Within Cloud Code and Codex, I have projects that are basically organized around their shared context in the form of repositories, and then agents get spun up to do specific things with them, which in a session of Codex or Cloud Code is actually more akin to spinning up a new orchestration agent, given how many sub-agents they might spin up to do whatever work you've given them. There's not a persistent named anthropomorphized agent. There's a persistent bundle of context that I can drop any set of agents into at any
19:29given time. Although it's not perfect, I think in some ways, a better analogy than teammates is something like on-demand consulting firms. Think about how an engagement works. You bring a goal, that's a prompt or a signal. The engagement lead or coordinator staffs it, that's an orchestrator. You never manage the staff, and you may never even meet them. The deliverable comes back with evidence. You sign off and you own the result internally. And when the engagement ends, the agents and people who did the work are gone, but the work and the records of the work stay. And theoretically, if you've done that well, the firm's methodology can get better every time. So not totally perfect, but I think directionally closer if we are plumbing
20:09for analogies that help people wrap their heads around how to use these new systems. Interestingly, a study published in the Harvard Business Review in May found some actual downsides to treating agents as teammates as well. The study looked at researchers working with BCG. Managers reviewed the same documents with errors planted in them. The study looked at researchers working with BCG. Managers reviewed the same documents with errors planted in them. The only difference was the label on who wrote them. An AI tool, a human colleague named Alex, or an AI team member named Alex3. When the work came from an AI teammate, Alex3, managers checked it less carefully. And the clearest effect was that companies that said they already treat AI as employees, where scrutiny dropped about 16%. One manager describing an agent named
20:49Kevin pointed out that a problem was, when Kevin, aka the AI team member, makes a mistake, Kevin literally can't be accountable. It's always important to not overstate this. It was one study. It happened very early on in the adoption of agents. And to be clear, if you are finding success in your company using the teammate analogy, far be it for me to suggest that you shouldn't be doing so. One last note, you might be wondering if it is true that the agent-teammate paradigm, the named agent paradigm, isn't really how things are playing out. How does that interact with the personal agent paradigm where meta and chat GPT are having you name your muse or dot, doesn't that somehow sit in contrast to that, especially because those products are doing
21:30well? I don't necessarily think it does for the simple reason that comes back to the idea that focusing your interactions on the orchestrator is where you need to focus. Effectively, these personal agents are like the er-orchestrators. They have, theoretically, all of the context about everything in our lives, and so are going to be in the best position to orchestrate whatever agents and sub-agents we need to do anything, while leaving us to still only interact with one point of contact. Now, I think it's still too early to know if these two trend lines will fully merge, or if they just represent different parts of the AI experience, but I don't think they're as mutually exclusive as they might seem at first. Anyways guys, based on a
22:09close reading of lots and lots of recent shares, that is 7 things AI native companies do. And that is going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.


