The AI Engineering Skills Map for Knowledge Workers

Transforming knowledge work with agents

0:00Knowledge work is being totally transformed by agents right now. After years of promise, agents are actually here, and they are changing the way that knowledge work gets done. Broadly speaking, we are increasingly moving from doing our work to managing agents that do our work. But in that transition, what are the key skills that matter? Five that stand out to me are 1. AI capability mapping, 2. Context and harness management, 3. Problem and product prototyping, 4. New opportunity identification, and 5. Rapid new skill acquisition. When you combine these with a foundation of domain judgment,

0:34you get a type of knowledge worker that is more capable and more powerful than ever before. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:50All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Section, Blitzy, and Hyperagent. To get an ad-free version of the show, go to patreon.com slash aidailybrief, or you can subscribe on Apple Podcasts. And to learn more about sponsoring the show, send us a note at sponsors at aidailybrief.ai.

Cursor launches Origin repo hosting

1:10In the world of AI, nothing is safe as everything gets rebuilt. The latest reminder of this comes from Cursor, who are taking on GitHub with their new repo hosting platform, Origin. The pitch is pretty straightforward. Take Git hosting and add better integrations for the coding agents you're already using. Origin allows developers and their agents to access the codebase from the same surface they're already working in. That makes it easier to run natural language queries across the codebase or use agents to handle comments and commits without context switching or using connectors. Still, maybe their biggest pitch is around platform stability. GitHub service has been widely perceived to be

1:45degrading over the past year, with frequent outages and issues. As if on cue, right around the time that Origin was announced, GitHub experienced a six-hour service degradation. Matt Palmer, a dev experience staffer at SpaceX AI, nodded to the issue, commenting, We were going to ship this earlier, but GitHub was down. Vercel CEO Guillermo Rauch piled on, saying, You can now host your repos in Cursor Origin and deploy to Vercel via Cursor Origin, which is itself hosted on Vercel. And unlike GitHub, it's online. Now at this point, complaining about GitHub is basically just part of the experience of being a

2:20developer or, increasingly, being a not-developer who works with code thanks to coding agents. Still, as we've seen so many times in the past, major switches like this ask a huge amount of users. And it's not clear whether a promise of better platform stability and AI integration is enough to drive users to switch. GitHub has incredibly strong lock-in, given how painful migrating codebase infrastructure is. And while there were plenty of people ready to immediately crown Origin as a GitHub killer, many questioned the premise. Scott Dolinsky, a content creator

2:53focused on DevTools, wrote, I'm going to be real. GitHub's outages are super annoying. I just don't see myself moving all my stuff to Origin. Maybe some of y'all moving your stuff there will take the pressure off GitHub's servers for me, though. Kush, a software developer at Cursor, used the opportunity to highlight one of the key design decisions with Origin, commenting, We get it. Moving your code is extremely hard. That's why Origin supports mirroring. Leave your code where it is and let us sync it, so you can always access it performantly through Cursor, agents, pull requests, local, etc. And if you end up liking Origin, you can make the decision to

3:24detach from GitHub later on. This means, in other words, that GitHub can remain the developer's system of record, theoretically minimizing the risk of breaking production workflows. Others weren't sure that Cursor was really in a position to actually promise better long-term platform stability. Veteran developer Michael Kove wrote, I don't trust them. Not that Cursor is a bad company. They are not. But that they don't have a proven track record of security and infra OPSEC. Because, obviously, it's a new offering. I like Cursor as a coding harness. I think they have a truly innovative product, and I do trust that my code is not going to be used in training.

3:56But a code repository store? That's a bit of a different concern. Ultimately, where Kove landed was to say that he would be happy trying out Origin for personal projects, but that he's going to need to see a longer track record before moving customer projects over. The launch itself is happening as an early preview to existing customers, and some were disappointed about that decision. Still, most felt that the inevitable early teething issues made that call completely understandable. Still, the promise for this kind of product is obvious, as software organizations integrate agents deeper into how they work. One developer noted that Origin makes it easier to connect

4:28automations to your codebase, for example, triggering an agent on a code change or running scheduled tasks from a single platform. And this idea of a GitHub that is built from the ground up, assuming agents are pushing most of the code, is what has people interested. I think, if nothing else, the response shows that there is demand for a new agent-first approach to codebase management. But that does not make Origin's path any easier, as there is also a very high bar for ripping out and replacing core infrastructure like GitHub. One has to be pretty impressed, though, with the absolute velocity of cursor shipping, especially given that this is happening in the context

5:00of the SpaceX acquisition being complete. Now, it's early days, but stereotypically, acquisitions tend to significantly slow down the rate of innovation from the acquired companies. SpaceX AI seems determined for that not to be the case, and in their announcement blog, confirming that the acquisition is complete, Cursor said that Grok 4.6 was an early look at what the two companies plan to build together. Perhaps framing what we can expect from the division of labor within SpaceX AI, Cursor wrote, Now, speaking of companies pushing at an incredible rate, more details have emerged about Anthropic's

Anthropic reports massive revenue growth

5:40exact financial situation, as an official run rate figure of $65 billion is leaked. According to sources who have seen the information, Anthropic hit $65 billion in revenue run rate at the end of July. The figures were reportedly shared with investors as part of a regular update. This represents a seven-fold increase since the beginning of the year, and a roughly 40% jump since Anthropic disclosed a figure of $47 billion in May. Now, obviously, what makes their growth rate so important right now are the astronomical numbers the company is seeking as a valuation for their IPO. Reports from last week suggested that some investors expect the valuation to come in at $2 trillion,

6:13and some quoted an 800% growth rate as justifying an even larger number. These figures would put the annualized growth rate at 550%, still incredibly strong, but slowing compared to the first half of the year. Now, believe it or not, because the sheer growth rate behind these numbers wasn't as stratospheric as it was earlier in the year, there were actually some investors who were a little bummed out by this. Author Take Him wrote, Come on, people. You can't extrapolate month-over-month growth acceleration to infinity. This is still insane growth. And what's more, when you combine this with the $40 billion in annualized

6:45revenue run rate that OpenAI CFO Sarah Fryer has been reporting, that means that between these two companies, you're talking about $100 billion in revenue run rate, up from low single-digit billions at this time last year. Lastly today, we got official information about the Stripe OpenRouter

Stripe acquires OpenRouter

7:00deal. The deal appears to be complete, and Stripe will acquire OpenRouter for $7 billion. That price tag is, of course, less than the rumored $10 billion, but still a huge markup for OpenRouter, who last raised funds at $1.3 billion in May. Indeed, one of the big strands of conversation is the idea that Stripe overpaid. After the news broke, there were infinite variations on this post from Vercel's Max Later, who wrote, Big congrats to the OpenRouter folks, but can someone explain how an AI gateway justifies $7 billion? Surely Stripe can make their own gateway, and it's not like switching costs for users are particularly high. Fintech engineer Samir disagreed. He wrote,

7:34The reported Stripe OpenRouter price is inexpensive. Everyone calling it that is anchoring on the wrong number. $7 billion only looks crazy if you judge it against what OpenRouter is worth on its own, and that's not the number that matters. What matters is what the deal does to the acquirer's value. Look at what Facebook did with WhatsApp or SpaceX with Cursor. Big acquirers pay up, and they pay up relative to their own market cap, not the target. So here's the real test. Ask yourself if you actually think buying OpenRouter for $7 billion improves Stripe's own valuation prospects by more than 5%. If you believe it does, the price makes total sense. It pays for

8:08itself. Then it's a competitive tension question. What do you pay to make sure nobody else gets the deal? You get a little more generous than the hurdle. That's how $5 billion becomes $7 billion. And that's why $7 billion makes sense to me. To be honest, I think everyone might be overthinking this a little bit. Stripe is and views itself as the core financial plumbing for the internet economy. Tokens are a new essential currency in that economy. Recently, people have realized that not all tokens are created equal, and they need to be able to easily move between different categories of tokens for different types of use cases. And OpenRouter is the leading company doing that.

8:40To get all those customers, all that infrastructure, the brand, and to have it come online and into your system right away, rather than later down the line after you've introduced your own version and ask people to switch over, especially in the context of the speed at which AI is moving. Basically, you just pay the price that it takes to get the deal done. And it turns out that was $7 billion. A great outcome for the OpenRouter team, of course, but my guess is that this pays off for Stripe in big ways as well. For now though, that's going to do it for the AI Daily Brief Headlines Edition, next up the main episode. A new study from KPMG in the University of Texas

9:15at 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, evaluating, 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. Here's a harsh truth. Your company is probably spending

9:50thousands or millions of dollars on AI tools that are being massively underutilized. Half of companies have AI tools, but only 12% use them for business value. Most employees are still using AI to summarize meeting notes. If you're the one responsible for AI adoption at your company, you need Section. Section is a platform that helps you manage AI transformation across your entire organization. It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value. The result? You go from rolling out tools to

10:20driving measurable AI value. Your employees move from meeting summaries to solving actual business problems. And you can prove the ROI. Stop guessing if your AI investment is working. Check out Section at sectionai.com. That's S-E-C-T-I-O-N-A-I dot com.

10:38Blitzy deeply understands your codebase before it writes code. Here's the first place that pays off. Security in the age of AI. Vulnerabilities don't live in isolation. They live buried inside millions of lines of interconnected code, where patching one thing quietly breaks three others. That's why surface-level scans fail. Blitzy starts from its knowledge graph of your entire application, identifies and surfaces CVEs across the full estate, proactively recommends patches, and can execute the PR. Each fix is grounded in how your systems connect and validate so nothing new breaks. And the knowledge graph dynamically updates, keeping you ahead of an ever-accelerating threat landscape. One Blitzy

11:10customer resolved 21 active CVEs across six core microservices in four days. Zero compile errors, every validation scan clean, months of planned work fixed in less than a week. Security remediation grounded in real architectural context at the speed of compute. Harden your codebase at Blitzy.com. That's B-L-I-T-Z-Y dot com. This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together. New users get $1,000 in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. HyperAgent deploys always-on agents in the cloud,

11:45doing real work across the tools your team already uses. Marketing's agent turns competitor moves into landing pages. Sales's agent enriches leads, drafts emails, and updates the CRM. Ops agent chases the paperwork and tracks the budget. Every agent has access to shared context and follows your rules about scope and approvals. It's time you add agents that feel like teammates. Hire yours at HyperAgent, built by the team at Airtable. Claim your $1,000 in inference at hyperagent.com slash AI Daily Brief.

Five AI skills for knowledge workers

12:10Welcome back to the AI Daily Brief. Today, I've got a fun one for you. It's very clear that the skills of knowledge work are changing and changing fast. Not that everything's abundant, but all of a sudden, knowledge workers have these totally new capabilities brought on by new tools and new ways of working. And of course, alongside that comes with the challenge of figuring out how to harness all that power. Today, we're going to go through my AI Engineering Skills Map for Knowledge Workers.

12:41It's a list of five skills that I think broadly are now relevant for knowledge workers of all stripes who want to work in this new paradigm. And the inspiration for this came from Coursera founder Andrew Ng's AI Engineering Skills Map, a post that he dropped on X last week. Now, in his post, he is talking about AI engineering very specifically in the context of software engineering. Andrew writes, AI allows us to build software very differently today than in 2022, and everyone with the skills to take advantage of this shift has numerous exciting project and job opportunities. But with

13:13the noisy, hype-filled information environment around AI, what are the most valuable skills for you to learn? I've been working with my team to synthesize a map of AI engineering skills in order to help developers prioritize what to learn and help employers hire skilled developers. For Andrew, the four most important AI engineering skills are one, building and deploying AI applications. 2. Software engineering fundamentals. 3. Using coding agents. And 4. Shaping the build. I don't want to spend too much time on these before getting to our interpolation, but to give you a

13:44quick flavor, on something like software engineering fundamentals, he's basically saying that even if you are using coding agents, as is the third skill, you need to deeply understand how software works, i.e. understanding trade-offs between cost, scalability, reliability, speed, and more. Even if you are not writing the code directly, this is going to lead, as he puts it, to better decisions in choosing your software stack, designing systems architecture, designing your data store, testing, and so on. So foundations and fundamentals still matter. However, there are obviously new skills as well. As Andrew points out fairly uncontroversially at this point, using

14:16agentic coding effectively is now a key skill for every developer. And it is indeed a skill. It involves understanding limitations, how to steer them, how much to intervene, how much to leave them alone, etc. So these are the types of new skills for AI engineering. But what's interesting is that AI engineering is infiltrating the rest of knowledge work as well. And here, admittedly, we're using AI engineering in sort of two ways. We're using it in the metaphorical sense of engineering AI to use AI well, but also the literal sense of engineering-style skills coming to non-software engineering fields.

14:50As you'll see, as many of you have experienced, even without pretending that somehow knowledge workers outside of software engineering are all of a sudden going to be software engineers, the capability that knowledge workers have now to build and use code to solve problems and create opportunities is creating a dramatic shift in how we all work. Still, while I'm going to get into these five new skills of AI engineering for knowledge workers, I do also want to suggest that a foundation is and remains domain judgment. And the analogy here is certainly the software

15:21fundamentals from Andrew's post. It's things like the ability to define quality, recognize trade-offs, understand consequences, take responsibility for decisions. It is the judgment that comes with experience in that field. In other words, just because ChatGPT can write all the copy and make all the ad assets, we wouldn't expect someone with no experience in marketing to be able to plan and execute a marketing campaign really well because they lack that domain judgment. And I think it's important to note in the context of your specific work that domain judgment is not just general.

15:53It's not, to continue that example, marketing in general that matters. It is marketing in the context of your organization. In every field, there are going to be things that haven't or even can't get encoded into AI training data and that shape the difference between work that is passable, good or great. Now, one of the interesting challenges that the continued need for domain judgment creates is the apprenticeship question of how young workers can develop that domain judgment if more experienced workers with the domain judgment are simply using the tools to do what

16:25the younger workers might previously have done. At the end of the episode, I'll get into a new organizational structure, which I think might be an interesting answer to that. But I do also want to point out that domain judgment doesn't exclusively come from 10 or 20 years of working in a particular firm or field. There are lots of different types of judgment that matter. Personal judgment, standards, pattern recognition, and trade-off awareness that you've already internalized. Borrowed judgment, expertise supplied through collaboration, review, explanation, and shared decision-making, i.e. younger workers can work with older workers. And of course,

16:56there's embedded judgment, which is expertise that's captured in examples, rubrics, policies, evaluations, the sort of context that you feed AI. But wherever the domain judgment comes from, it is the foundation for the rest of it, and does not suddenly leave when you introduce AI. But let's move on to these five skills. These are not somehow completely disconnected from one another. Indeed, each capability strengthens the next. And I would argue that it starts with AI capability mapping. One term that you'll frequently hear, popularized by people like Professor Ethan Mollick, is the idea of AI having a jagged frontier. What he and others mean by that

17:29is that while AI can absolutely blow you away in one minute, it can make a mistake the next minute you wouldn't expect the least capable intern in the organization to make. Capability mapping is in part understanding that jaggedness. It's about being able to know what AI is natively good at and what it isn't, where in the context of where it isn't great yet, how you can help it. It's the ability to match different tasks with different approaches to AI. In other words, understanding the difference between an assisted approach, a workflow automation, or a truly agentic solution.

18:01Increasingly, capability mapping is also about understanding which different models and effort levels are sufficient for different types of tasks. And capability mapping is about understanding how much human oversight is going to be needed. And unfortunately, and this will be a recurring theme throughout this, this is not something that in general someone can just hand you to learn. You kind of have to learn this by doing it yourself through trial and error. Which isn't to say that standards and norms and conventional wisdom can't be helpful guides on this process, but different people's experience with AI is going to be different. One person may find that on every

18:33different type of writing task they've had, Fable is by far the better writer. But for a particular type of technical writing, that is the writing that matters for you, 5-6-Soul blows it out of the water. Capability mapping is a combination then of broadly what we all know, and specifically what matters for your particular function or role. The next skill after capability mapping is context and harness management. Basically, how you set the AI up for success. We've talked a lot over the last calendar year about context engineering and more recently, harness engineering. Context engineering

19:05or context management is all about making sure the AI has access to the information it needs. Going back to that marketing example, there is likely to be a wild difference in how the AI performs for you if you provide it with past campaign performance, analytics, subjective reviews, customer feedback, as opposed to just giving it a prompt and hoping that it comes up with something great. Harness management builds on context to give us more ways to help the AI succeed. You can think about the harness as everything that surrounds the model that isn't just the context,

19:35meaning instructions documents, tool access, permissions, memory, and more. Now in both of these cases, some amount of this is going to be governed by your organization, but a lot of it is going to be customized by each individual worker and how well they've set up their particular instance of a model to be successful or not. Here we also start to understand that these skills are extremely dynamic. Best practices in context and harness management today will almost inevitably have changed six months from now. There might still be strong core ideas, but different models, different harness software are going to change

20:09the best practices for what you have to do to maximize the AI's environment. The next skill I'm calling problem and product prototyping. And this is my sort of catch-all for the new capability that knowledge workers have to use building software and pushing code that used to work very differently. And I want to be clear here, and this is why I use the term prototyping, although it's not perfectly accurate, that what I am not talking about here is all of a sudden marketers being the software engineers. But it turns out that when everyone can use code

20:39in a much more robust way, a lot of previous work can be done much more effectively. Maybe the most obvious example that is going to resonate across a broad cross-section of you listeners is anything having to do with analytics or data. Continuing our analogy of the marketer, let's assume that there's some regular reporting interval for figuring out how well a campaign is doing. Previously, getting all that information ready was probably a ton of work going to each of the different platforms' analytics suites, downloading what you had access to, taking it and moving it into another system like Google Sheets or Excel, running analysis on it, taking that analysis from

21:14one channel and trying to combine it with other channels, and then once you actually came to some conclusions, turning that all into presentations that could help share it with other parts of the organization. That changes fairly dramatically in a world where marketers can push code. All of a sudden, they can build not only an internal dashboard, but the engine underneath it that is automatically ingesting and processing that information in a totally different way. Instead of having to manually go and get all that data, they can connect that marketing analytics engine to the APIs of

21:44those services and ingest that automatically. The first layer of analysis can come not from them crunching through Excel, but by the AI surfacing interesting insights. Of course, where domain judgment comes in is that there is an additional translation layer that's still required to turn that into real actionable insight, but all of a sudden a huge amount of time is freed up to ask those sort of judgment-style questions and do that sort of judgment-style work. I truly do believe that even though knowledge workers are not and should not think of themselves as turning into full product

22:15managers and software engineers, being able to build things to solve problems and do parts of our jobs is perhaps the most significant shift in how knowledge work will happen that we've ever experienced. For those of you who have been on the fence, it is worth taking the time to go messily clunk your way through codecs and cloud code, which parts of your work could be transformed if you were building things that did the work instead of doing it yourself. Skill 4 is sort of the matched pair and advanced level of Skill 3. If Skill 3 is all about using the new capability to build software to solve

22:49problems of the existing way you work, Skill 4 is about identifying the new opportunities that building software and pushing code open up in what your job actually consists of. Instead of asking, how can AI help with different parts of my job? It's about asking, what can we do now that was previously impossible or uneconomic? In other places, I've referred to the idea of the infinite backlog. The idea that every knowledge worker has this endless list of things that they would do if time and resources were no limit. Agents bring forward that infinite backlog and make a much bigger

23:22portion of it actually viable. And while I don't think that there's any super general shorthand for how to identify the new opportunities of things that were previously impossible that you should be doing now, one small tip that I have seen help is to imagine that your organization gave you access to a team of software engineers and said, you can do whatever you want with these folks. What would you do then? Maybe the first layer would be things like we saw in number three, where you take a lot of the work that used to be manual and turn it into actual automated or agentic systems. But before long, you might start to run into totally net new ideas. I think we're going to start to see some weird and very cool

23:57things like marketing teams from small companies building and releasing games as part of their top of funnel. And honestly, if you've spent any time around this show, you'll know that I think this skill is where some of the most exciting outcomes of AI are going to come from. Skill five is a meta skill that cuts across all of these things, which is the ability to rapidly acquire new skills. The speed at which AI changes the opportunity set is incredibly quick. And while there is organizational and institutional inertia that will limit how fast new systems and processes can be

24:28integrated into the way that the organization works as a whole, there is far less inertia in how quickly you are able to integrate new ways in how you accomplish what you do. And in fact, your ability and your team's ability to integrate those new skills and new capabilities more rapidly probably has net positive impacts on how fast your organization can move as well. New skill acquisition is a combination of being able to recognize what new or adjacent skills and capabilities have suddenly become valuable. It's the ability to create the right type of space to actually go experiment and learn by

25:00doing and applying it to real world environments. And it's about the ability to assess the output and understand whether it's actually worth integrating into how you work. It is both mindset and discipline. Skill acquisition becomes not the work of semi-regular upskilling seminars, but a continuous process that happens in an ongoing way. So those are my five AI engineering skills for knowledge workers. Again, to recap, it's AI capability mapping, context and harness management, problem and product prototyping, new opportunity identification, and rapid new skill

25:31acquisition. I'm super interested to release this show and see which of these resonate the most with you, which ones you're struggling with. Obviously, I'm thinking a lot these days about how to support all of this. So your commentary and ideas are very welcome. Lastly, one flag for something that I'm thinking about for a future podcast exploration. I think this question of what happens to the next generation of domain judgment if young workers aren't being given the chance to develop that judgment because more experienced workers are simply using AI to do what those younger workers might have previously done is a really fascinating one. And one of the answers that I think is

26:04interesting to explore is a shift in thinking about AI as single player to AI as multiplayer, with a core unit of AI moving outside the individual and to the small team. I think that that could create some really interesting new opportunities, but that is a subject for a future show. For now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace!

26:39you