Angela Jiang | Training Data Summary

1. Key Themes

The Three-Layer Abstraction Stack: Knowledge → Execution → Coordination

Anthropic's platform roadmap is explicitly structured around three ascending layers. The knowledge layer covers model access and context primitives; the execution layer covers managed infrastructure and harnesses for agentic work; and the emerging coordination layer introduces "strategies" — meta-harnesses that assign different jobs to different tokens. This is the clearest public articulation of where Anthropic's developer platform is headed.

"You have knowledge and you have execution, you have coordination. And at the coordination layer, we're beginning to think of these things called strategies, where basically it's almost like a meta harness... if tokens aren't really fungible and you need to give them different jobs, like maybe this token is advising versus this token is executing, this token is dreaming versus this token is executing." — Lauren Reeder 00:11:03

Token Jobs as the New Unit of AI Optimization

The team has moved beyond thinking about models or model sizes as the primary lever and now frames optimization around what job each token is performing. This reframes the entire field of harness engineering.

"The interesting innovation is going to come more at that higher level on like the meta level. You can take any given token and spend that token on just executing or you could take that same token and choose to actually reflect on your past agentic sessions and write learnings to memory so that the next agent does a good job. Or you could take that token and advise with a bigger model so that a smaller model can execute and do a better job. Or you can say execute, execute. And then like a grader comes in is like, did you do a good job? No, you didn't, try again." — Katelyn Lesse 00:27:44

Harness Steering Is Obsolete; Strategy Harnesses Are the Future

Models have become so steerable that old scaffolding designed to keep them on track can now simply be deleted. The harness's new job is to let agents run longer and reason at a higher level, not to constrain their direction.

"If you look like two years ago, a lot of the harness was like a scaffold to kind of like tell the model to go from point A to point B... And now the models are actually very, very steerable. And so a lot of that steering, you could just put in the prompt... a lot of if you have harnesses that are like designed to kind of do that kind of like steering, you can delete that part — that part we actually frequently encourage you to delete." — Lauren Reeder 00:28:53

Open Ecosystem Over Walled Garden as Deliberate Strategic Posture

Anthropic explicitly does not care whether agents run on their infrastructure versus partners'. They partnered with Modal, Vercel, Cloudflare, and Amazon micro VMs for sandboxes rather than locking users into proprietary compute. The strategy is to own the architectural opinions and interfaces, not the infrastructure.

"Whether it runs on our infrastructure versus somebody else's infrastructure is actually not important to us, because the thing that's important to us is more that the architecture of how you put together these agents in a way that will be powerful in a way that will be reliable and scalable. We have strong opinions on that, and you can just conform to the interfaces that we put out there and plug those things in." — Katelyn Lesse 00:16:50

Manufacturing Is the Next Frontier for AI Adoption

Beyond the well-known coding and knowledge-worker use cases, manufacturing is emerging as a major new category. This was mentioned as a concrete, current trend — not a speculation.

"More recently, we're starting to see manufacturing really pick up as just a category where people are building with AI. And like one of our PMs, like getting on a flight to Detroit to go, like, figure out what these customers, like what they need and what's going on." — Katelyn Lesse 00:37:46

Healthcare's Legacy Systems Problem Is Spawning Unexpected Innovation

Healthcare companies operating on systems with no APIs are using computer use and creative automation workarounds — including literally running laptops to auto-generate interface specs — creating a new category of connectivity innovation.

"A lot of like healthcare companies that we kind of engage with... they're like the systems I'm working with, they don't even have APIs. Like that's a dream. And so, you know, how can they use computer use and things like this to be able to start to kind of automate and create more connectivity with our systems?... taking a laptop and trying to like run a bunch of things on it to auto generate a bunch of things that then their agents can go and use." — Lauren Reeder 00:36:09

Context and Connectivity as the True Innovation Layer Right Now

Across their most advanced customers, the highest-value innovation is not in model selection or prompt engineering but in how companies architect context retrieval, permission management, and connectivity across disparate systems.

"The general theme I would just give you is like, interestingly, a lot of the innovation that's most exciting out there right now has been this kind of like context and connectivity layer, which has been really fascinating." — Lauren Reeder 00:36:38

Token-Hungry Use Cases Drive Platform Strategy

Anthropic explicitly filters which verticals to invest in based on whether completing a task makes users want to do more — i.e., whether the task is inherently token-hungry and iterative, not one-and-done.

"We like industries where... the answer to that question, you say, I want to do more of that thing. So coding is obviously the one that we all know. And the great thing about coding is that what it's actually doing is that like, once you've finished a turn, you look at that, and you're like, that was incredible. I'm like unlocked, I'm going to do like more." — Lauren Reeder 00:20:22

Claude Tag Is an Iceberg Product — the Interface Is the Least Important Part

The perception that Claude Tag is "just a Slack bot" fundamentally misses the point. The value is in the context engineering, proactivity logic, and org-level harness underneath. The interface is deliberately interchangeable.

"The important part is all the kind of like context engineering and like architecture that we put underneath the hood... I think André Karpathy said it really well. It's like, it's like an org level harness. There's a lot of like complexity baked into that." — Lauren Reeder 00:23:48

2. Contrarian Perspectives

Harness Ownership Is Overrated for Most Use Cases

The prevailing view in the developer community is that owning your harness is critical for AI product differentiation. Anthropic's team argues most companies are optimizing the wrong layer and should let that go.

"The stuff that can be pretty generic and like less interesting to own and deal with is... getting your prompt caching right. Right. Like maybe that is not the world's most interesting thing. Choosing to clear out old tool calls from the context window and things like that. Right. Or like maybe a little bit less interesting. And you like go a layer higher into some of the stuff Angela's talking about." — Katelyn Lesse 00:32:47

Context Engineering Is Overdone as a Differentiator

While the AI industry obsesses over RAG pipelines and context window optimization, Anthropic's platform team thinks context is largely commoditized — any harness can handle context. The real differentiator is verification logic and token budget strategy.

"The context bit is actually a little like overdone. Like, yes, you're going to like throw in context and like that's but any harness can actually handle a lot of context. And so that's just more like you have the data. And if you have the data, then obviously you're uniquely qualified to do something useful." — Lauren Reeder 00:31:59

Model Routing Across Providers Is a Dead End

The conventional wisdom is that model-agnostic agent frameworks are strategically superior because they offer flexibility. Anthropic believes harnesses should be tightly coupled to a model family, and Vercel's move to bundle the full harness+model together validates this.

"I think we started to see like Vercel just did this with their agent harness, for example. Like some of these players in the space like come up a layer of abstraction and say actually like plug in the whole harness and the whole agent that's tied to a model family, which makes a lot of sense." — Katelyn Lesse 00:41:36

Capping AI Spend Is the Wrong Response to Token Rationalization

As companies scrutinize AI costs, the instinct is to cap spending. Anthropic argues this is directionally wrong and kills compounding returns. The right answer is smarter strategies, not less usage.

"What we try to kind of encourage our customers is like you don't want to like stop the innovation. Like if you are getting returns on top of this, you are shipping faster than ever before... The thing that gets dangerous is when you're kind of just like here's a cap and you're stuck within your cap." — Lauren Reeder 00:39:24, Katelyn Lesse 00:42:47

Best-of-N Sampling Is Dramatically Underutilized and Underappreciated

The AI community discusses larger models and longer runs as the main levers. Anthropic's internal experiments show a third lever — best-of-N sampling — often delivers far more return than either, but is nearly impossible to productionize today.

"You actually have like a third lever and tends to actually do a lot more than you think it does, which is that actually if you were to like best of end the thing, it would like give you a lot more returns. But like just to be just saying those words are fine... to actually build that thing and put it into production... That's like really, really freaking hard." — Lauren Reeder 00:44:31

3. Companies Identified

Anthropic AI safety and research company, creator of Claude. The episode's central subject — building a developer platform and ecosystem around Claude's capabilities.

"Platform is both our externally facing APIs, our developer platform that people build on top of when they want to build applications and systems that access Claude's intelligence, as well as internally we run our product infrastructure." — Katelyn Lesse 00:01:28

Modal Cloud compute infrastructure provider. Named as a first-class partner for self-hosted sandboxes within Claude Managed Agents.

"We launched self hosted sandboxes, and we partnered with Modal and Vercel and Cloudflare and a bunch of other folks, even like Amazon's new micro VMs, to have a first class offering where you can go plug any of those things in." — Katelyn Lesse 00:16:22

Vercel Developer infrastructure platform. Cited both as a sandbox infrastructure partner and as an example of the trend toward bundling harnesses tightly with model families.

"We started to see like Vercel just did this with their agent harness, for example. Like some of these players in the space like come up a layer of abstraction and say actually like plug in the whole harness and the whole agent that's tied to a model family, which makes a lot of sense." — Katelyn Lesse 00:41:36

Cloudflare Network and cloud infrastructure provider. Named as a self-hosted sandbox partner.

"We partnered with Modal and Vercel and Cloudflare and a bunch of other folks, even like Amazon's new micro VMs." — Katelyn Lesse 00:16:22

Amazon Web Services (AWS) Cloud platform. Named both as a hyperscaler integration partner and a sandbox infrastructure partner via their micro VM offering.

"We really care about like bringing our platform really, really close to that business. This is why we spend a lot of time with the hyperscalers, integrating really closely directly with them like AWS, Google, so on and so forth." — Lauren Reeder 00:02:31

Google Named as a hyperscaler with which Anthropic integrates directly.

"We spend a lot of time with the hyperscalers, integrating really closely directly with them like AWS, Google, so on and so forth." — Lauren Reeder 00:02:31

Shopify E-commerce platform. Cited as an early pioneer of building an internal agentic platform (their "River" system), which informed the design of Claude Tag.

"We had been seeing people in the industry go and say, like Shopify did this with River, Square, Block recently did this with BuilderBot." — Katelyn Lesse 00:22:20

Block / Square Fintech company. Cited alongside Shopify for building an internal agentic platform called BuilderBot.

"Square, Block recently did this with BuilderBot. There's like a few of these examples where people said, I'm going to build like an agentic platform internal to my company." — Katelyn Lesse 00:22:20

Stripe Payments company. Mentioned as Katelyn Lesse's prior employer, used as context for how mature engineering organizations manage infrastructure costs (AWS bills) — framing the analogy for how companies should manage AI spend.

"Before working at Anthropic, I was at Stripe and we were kind of in the very reasonable era of like we paid a lot of attention to our AWS bill." — Katelyn Lesse 00:41:54

4. People Identified

Katelyn Lesse Head of Anthropic Platform (external/internal APIs and developer infrastructure). Previously at Stripe. One of the two primary guests and a key architect of Anthropic's developer ecosystem strategy.

"We found that we could piece together our primitives and stand up all the same infrastructure that we're finding ourselves standing up internally to power our own products and arrive at some higher order abstractions that let you do more agentic work out of the box." — Katelyn Lesse 00:06:40

Angela Jiang Product leader on Anthropic Platform. Co-guest, brings the investor/product lens to the conversation and asks incisive questions about ecosystem design.

"How do you think this all comes together into a broader ecosystem beyond just the things that you guys are building? How do you help support people building products on top of it?" — Angela Jiang 00:12:10

Andrej Karpathy AI researcher and former Tesla/OpenAI figure. Cited approvingly for his framing of Claude Tag's architectural complexity.

"I think Andrej Karpathy said it really well. It's like, it's like an org level harness. There's a lot of like complexity baked into that." — Lauren Reeder 00:24:44

5. Operating Insights

Prompt Caching Is a Mandatory Cost Lever, Not Optional

The Anthropic platform team is categorical about this: prompt caching is not a nice-to-have but a significant cost-reduction tool that nearly all builders should implement immediately.

"Best practices are just stuff like prompt caching, do it. You're going to save a lot of money and token costs. Obviously like try to keep your context window clear." — Katelyn Lesse 00:26:52

Dog Food Internally and Open Early Access Simultaneously to Avoid Over-Indexing

Building only for internal users or only for external users creates blind spots. The Anthropic team runs both in parallel to ensure primitives stay general-purpose.

"A lot of the time what we'll do is dog food something internally at the same time that we open up early access of some sort with external customers so that we can kind of get a range of feedback and bring those things back into the platform." — Katelyn Lesse 00:05:44

Use MCP as an Agent-to-Agent Connectivity Layer, Not Just Tool Integration

One customer insight that emerged: exposing an MCP server on top of one agent so a different agent (potentially on a different model/platform) can call it as a tool — enabling cross-platform, modular multi-agent architectures.

"What if I expose an MCP server on top of this agent so that I can then go and like have this other agent call a tool on that agent... have these things just be more modular and be able to work together... we sat down with them and worked through it and it worked perfectly and it was pretty cool." — Katelyn Lesse 00:36:49

The Right Response to AI Cost Pressure Is Smarter Strategies, Not Caps

When AI spend is scrutinized, the productive move is to identify whether the same outcome could be achieved more cheaply through better token strategies — not to impose blanket spending limits that kill innovation.

"There are a few different ways we probably could have accomplished that outcome, right? And one is like you take Opus and you run it all night and you do something crazy. And another is maybe to get a little bit smarter with the strategies that you put together in order to create that same outcome within a lower cost. And I think that's the like next layer of thinking that everyone's going to start to do." — Katelyn Lesse 00:43:15

Evals Are Non-Negotiable for Agentic Production Deployments

Mentioned almost as a confession — the team was surprised they got so far into the conversation before saying it. In agentic systems, evals are the minimum viable quality gate.

"Evals. I'm surprised we got this far into this thing before one of us said the word evals, but like you need evals to make sure that what you're trying to accomplish is performant." — Katelyn Lesse 00:27:18

6. Overlooked Insights

"Best-of-N" Production Deployment Is the Unrealized Alpha That Anthropic Is About to Commoditize

This was mentioned extremely briefly and almost glossed over, but it is potentially one of the most commercially significant signals in the episode. Anthropic's internal experiments show that best-of-N sampling (running an agent multiple times and selecting the best output) dramatically outperforms simply using a bigger model or running longer — but building this in production is currently so hard that almost no one does it. Anthropic is explicitly building this into their "strategies" layer to make it trivially easy. The team that productionizes best-of-N at scale will unlock a step-change in agent quality for the entire ecosystem — and Anthropic is planning to hand it out as a platform primitive. Any company that adopts this early will have a meaningful quality advantage before it becomes table stakes.

"You actually have like a third lever and tends to actually do a lot more than you think it does, which is that actually if you were to like best of end the thing, it would like give you a lot more returns. But like just to be just saying those words are fine... to actually build that thing and put it into production... That's like really, really freaking hard. And you end up building all these like custom harnesses... But we're seeing like this is where the alpha is." — Lauren Reeder 00:44:31

Shadow IT Is Already Driving Enterprise AI Adoption — and It's a Platform Distribution Signal

This was mentioned in passing as a cost management problem, but it is actually a profound distribution insight. Employees are self-procuring Claude Code and similar tools without IT approval, meaning bottom-up organic adoption is already happening at scale inside enterprises. This mirrors the Slack/Dropbox playbook from a decade ago and suggests that the winner in enterprise AI platform won't be determined by top-down procurement cycles — it will be ratified after the fact by whoever employees already chose on their own.

"The way that AI spend has erupted inside their company has been through some kind of like shadow IT. You know, like their employees just like want to use it. They find a way. They end up procuring it themselves. And before you know it, like half your org has like found some way to install Claude Code." — Lauren Reeder 00:39:24