"models have gotten really good at working for longer and they've gotten smarter and they're able to do more things and so a year ago we were still very much in this like human in the loop for all the things that you're doing with a model... and I think over the past year we've gotten more and more into this like models can work for a long time" - Katelyn Lesse [00:01:25]
"you do tackle the hardest problems whether they're societal economic technological seeing that kind of revolution through is exactly what I'm here to do" - Angela Jiang [00:10:00]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"from a product angle the kind of philosophy that we have with manage agents is that you should do as a builder on top of the platform you should do the things that differentiate you and you shouldn't have to do the things that are undifferentiated" - Angela Jiang [00:10:30]
"we have to be a team that's really ready to get punched in the face... everyone's got a plan until you get punched in the face and like the things that are punching us in the face are there's like really hard safety problems to solve or there's really hard infrastructure or scale problems to solve" - Katelyn Lesse [00:39:08]
"sonnet executing with opus advising ends up getting almost opus level performance and it's actually cheaper than just sonnet because opus taught it how to do its job better and it used less tokens to get the job done" - Katelyn Lesse [00:22:20]
"the most basic version of a harness is like a while loop that's literally just like go back and forth between like get input from the user ask the model what it thinks then call a tool over here and you keep that thing running" - Katelyn Lesse [00:23:54]
"you actually need to build like a portfolio very very quickly so that you can kind of almost like if any of the bets hit you kind of win and it's a really weird way to do a product cuz... it's more like you actually want the portfolio" - Angela Jiang [00:38:08]
Speakers & Credentials
Angela Jiang: Product Lead on the Claude Platform Team at Anthropic. Former Product Lead at OpenAI (working on the API product) and former Product Lead at Stripe.
Katelyn Lesse: Engineering Lead on the Claude Platform Team at Anthropic. Former Engineering Lead at Stripe working on Stripe Connect.
Lee Marie & Co-Host: Partners/Hosts at Kleiner Perkins (interview recorded for Builders Season 2 at Anthropic's San Francisco office).
1. Executive Summary
Shift from Prompting to Long-Running Agents: Over the past year, AI development has evolved from real-time "human-in-the-loop" prompting to autonomous, long-running agents capable of handling end-to-end knowledge work 00:01:27.
Infrastructure as the Primary Bottleneck: As models achieve longer execution horizon capabilities, the critical layer of differentiation has moved from steerability to underlying platform infrastructure, secure sandbox environments, and error-recovery mechanisms 00:02:30.
Differentiated vs. Undifferentiated Engineering: Anthropic's core product design thesis advocates that builders should focus exclusively on domain-specific customization while leveraging managed platforms (like Claude Managed Agents) to absorb the heavy lifting of distributed systems and harness mechanics 00:10:30.
Rethinking Token Economics & Model Roles: Tokens are not interchangeable commodities; strategic model delegation—such as assigning Claude 3.5 Sonnet as the executor and Claude 3 Opus as the advisor—yields higher intelligence output per dollar spent while consuming fewer total tokens 00:22:20.
High-Leverage Small Teams: Operating at immense scale with only ~200 people on the platform team, Anthropic relies on AI-assisted workflows to accelerate execution, requiring product managers and engineers to focus on root problem-solving rather than traditional administrative coordination 00:33:36.
[00:16:16] Security, Observability, and Enterprise Trust Boundaries
[00:17:31] Platform Stack Decisions: Self-Hosting vs. Anthropic Managed Services
[00:18:53] Myth Busting: Data Training Policies & Token Fungibility
[00:21:36] Strategic Token Utilization: Advising, Outcomes, Graders, and Dreaming
[00:23:33] Deconstructing the Agent Harness: Loops, MCP, and Meta-Harnesses
[00:26:02] Common Misconceptions & Redesigning Workflows for AI-First Execution
[00:27:54] UX Form Factors & Integrating Agents into Human Workspaces (Slack)
[00:30:36] Product vs. Forward Deployed Engineers (FDEs) in Enterprise Rollouts
[00:33:31] Managing a High-Leverage 200-Person Team at Massive Scale
[00:36:13] Pure-Play Product Management & Portfolio-Based Development
[00:39:03] Culture of Adaptability: "Getting Punched in the Face"
[00:40:26] Personal Dogfooding Workflows & Memory Systems
3. Detailed Thematic Summary
Paradigm Evolution: From Steerability to Long-Running Infrastructure
Shift to Long-Horizon Workflows: One year prior, state-of-the-art AI usage concentrated heavily on tight human-in-the-loop interactions and intense prompt engineering to force models toward single-step tasks 00:01:27. Current frontier models run continuously in the background, autonomously recovering from intermediate execution errors 00:02:30.
Shift in Technical Bottlenecks: Because baseline model intelligence can interpret high-level intent, engineering complexity has migrated away from context scaffolding toward background distributed systems, system scale, cost optimization, compliance, and multi-tenant security 00:01:55.
General Knowledge Work Automation: Complex enterprise domains such as corporate finance, legal analysis, and medical workflows are moving from micro-edits (e.g., updating a specific cell in an Excel sheet) to high-level outcomes (e.g., generating a full Discounted Cash Flow model and investment thesis autonomously) 00:03:23.
Platform Architecture & Ephemeral Sandboxing
Decoupled Architecture for Resilience: Building reliable long-running agents requires decoupling the durable agent state machine (the harness/brain running on persistent servers) from the execution sandbox environment where code runs 00:12:48.
Handling Ephemeral Sandbox Failures: Traditional container sandboxes are inherently ephemeral and prone to drops or crashes 00:12:36. By spawning isolated execution sandboxes strictly when tools need to execute, and destroying them immediately afterward, a sandbox failure will not crash the master agent state machine 00:12:59.
Differentiated vs. Undifferentiated Abstractions: Anthropic positions low-level harness stability, container management, and distributed execution as undifferentiated platform overhead, encouraging enterprises to spend engineering capacity exclusively on domain-specific prompts, custom tools, and context engineering 00:10:30.
Debunking the Fungibility Myth: Treating tokens as generic, interchangeable commodities like electricity or oil ignores how token quality, model sizing, and operational strategy directly impact total compute costs 00:21:24.
Advisor-Executor Model Topology: Running a smaller, faster model (e.g., Claude 3.5 Sonnet) to handle continuous execution while routing hard sub-problems to a larger, more capable model (e.g., Claude 3 Opus) yields near-Opus level intelligence at a lower cost than unassisted Sonnet execution, due to a drastic reduction in total token iterations 00:22:20.
Specialized Token Roles:
Graders: Deploying a secondary evaluator agent to judge outputs against a predefined rubric before releasing the result 00:22:39.
Dreaming: Running off-cycle background jobs that analyze historical execution logs to automatically write reusable skills and memory entries into the system context 00:23:02.
Redesigning vs. Patching Workflows: A common enterprise pitfall is shoehorning agents into fragile legacy human workflows (e.g., automating step-by-step KYC banking steps) 00:26:19. Successful implementations decompose legacy processes and redesign them from scratch to be natively agent-first 00:27:11.
Data Privacy Guarantees: Enterprise adoption concerns regarding proprietary data loss are addressed directly: standard customer data submitted via Anthropic APIs is explicitly excluded from model training datasets 00:19:05.
Observability & Trust Boundaries: Enterprise governance relies on multi-layer controls: enforcing security parameters via Bring-Your-Own-Sandbox (BYOS) architectures alongside full execution visibility for auditing and compliance verification 00:16:30.
Team Design, Culture, and Product Management in the AI Era
Hyper-Leveraged Small Teams: Anthropic’s platform team manages global infrastructure handling millions of users and billions in revenue with approximately 200 people 00:33:31.
Pure-Play Product Management: As routine administrative and coordination tasks become automated by internal Claude agents, Product Managers must focus purely on root-cause problem identification, domain positioning, and high-level hypothesis generation 00:36:28.
Portfolio-Based Development: Given rapid underlying model advancements and shifting execution realities, teams must manage product initiatives like an investment portfolio—launching multiple parallel bets with high failure tolerance rather than betting sequentially on rigid roadmaps 00:38:08.
The "Get Punched in the Face" Cultural Doctrine: Internal engineering culture demands immediate adaptation when unexpected platform load, safety challenges, or scale bottlenecks alter quarterly plans overnight 00:39:08.
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Anthropic Platform Team Size
~200 people
Total headcount managing platform product, infrastructure, and scaling global Claude API traffic
Rather than relying on a single model for an entire workflow, task topology is split between an Executor (a fast, cost-effective model like Claude 3.5 Sonnet) and an Advisor (a high-reasoning model like Claude 3 Opus). The Executor conducts standard tool calls and routine processing, escalating to the Advisor only when encountering high-ambiguity edge cases. The strategic irony of this framework is that introducing a more expensive model can actually reduce the total cost of a run. By eliminating circular error loops and hallucinated retry cycles early, the system consumes fewer overall tokens to achieve a successful output.
To prevent catastrophic state loss during long-running tasks, the agent state machine must be structurally isolated from the code execution environment. The harness (holding system memory, prompt history, and operational rules) runs on persistent host infrastructure, while code execution is offloaded to ephemeral, isolated sandboxes spawned strictly per command. If a sandbox environment experiences an unhandled exception, resource crash, or network timeout, the main agent state remains intact, allowing it to spin up a clean sandbox and resume progress cleanly.
When the capability baseline of underlying technology advances exponentially every few months, traditional linear software roadmaps (e.g., 6-to-12-month sequential feature releases) become counterproductive. Product managers must instead operate like venture capitalists—maintaining a broad portfolio of parallel experiments. Because feature development velocity is compressed by AI leverage, the cost of initiating an experiment drops dramatically. Teams can afford high failure rates across individual features, provided the surviving bets align with emerging model capabilities.
As generative tools absorb administrative overhead—such as writing ticket descriptions, basic project tracking, status updates, and boilerplate code generation—the boundary of human value shifts to the fundamental core of each discipline. For product managers, this means spending less time on project management and more time evaluating user psychology, market positioning, and core product thesis. For engineers, manual syntax authoring gives way to system design, threat modeling, and state-machine verification.
The Story: While working at Stripe, Angela Jiang and Katelyn Lesse spent months designing the high-stakes "V2 Accounts API"—a legacy-defining abstraction built to last for decades. During an early user research session, a developer ingested the freshly written documentation, fed it directly into Cursor, and issued a simple prompt: "Integrate this." The model auto-generated ~70% of the functional integration code in seconds.
Why It Was Told: Angela highlighted this specific moment as her personal catalyst for leaving traditional developer tooling to join frontier AI labs, realizing that long-form knowledge work and developer workflows had permanently changed.
The Story: Upon departing Stripe around the same time, Katelyn Lesse joined Anthropic while Angela Jiang joined OpenAI. Over the following months, the two former colleagues routinely met up, each attempting to convince the other to jump ship to their respective organization—Katelyn pitching the platform needs at Anthropic and Angela advocating for OpenAI. Ultimately, Katelyn won the debate, leading to Angela joining Anthropic's Claude Platform team.
Why It Was Told: The anecdote demonstrates the tight talent density and competitive dynamics across AI labs, as well as the close professional alignment required between engineering and product leadership.
The Story: Katelyn Lesse shared a personal anecdote from a trip home to New York during the holidays. When explaining her work to family members who previously struggled to understand specialized software platform engineering, they instantly recognized Claude and expressed enthusiasm about using it in their daily lives.
Why It Was Told: This story highlights how rapidly AI tools transitioned from obscure tech-industry infrastructure into widely recognized consumer household names, shifting the underlying scaling challenges overnight.
7. References & Recommendations
Companies & Platforms
Anthropic – AI research lab and creator of the Claude model family and developer platform [00:00:51].
OpenAI – AI research laboratory; previous employer of Angela Jiang [00:07:55].
Stripe – Financial infrastructure and payments company; former employer of both Katelyn Lesse and Angela Jiang [00:04:54].
Kleiner Perkins – Venture capital firm producing the Builders video podcast series [00:00:31].
Cursor – AI-first code editor referenced during the Stripe API integration story [00:08:52].
Slack – Enterprise communication platform host to native Claude integrations [00:29:07].
Technical Protocols, Products & Concepts
Claude Managed Agents – Anthropic’s managed abstraction platform handling background infrastructure for long-running workflows [00:14:08].
Model Context Protocol (MCP) – Open protocol used for connecting AI models safely to external data sources and local tools [00:16:00].
Messages API – Anthropic's core low-level API primitive exposing raw model token streaming [00:18:02].
Claude 3.5 Sonnet / Claude 3 Opus – Specific model tiers within Anthropic's lineup discussed in multi-model routing strategies [00:22:20].
Stripe Connect & V2 Accounts API – Complex financial platform architectures referenced as background examples for system design [00:04:58].
Sep 3, 2026
EP. 20 | 10,000 Pitches, 55,000 Deals, and AI: How Pranav Pai Finds Winning Startups | 2 Sept 2026 | Clearing The BLUR
"My only regret since '91 is we have not become antisocialist." Pranav Pai 00:04 http://www.youtube.com/watch?v=lPkjuRztHcg&t=0m4s "The assumptions the government makes—every Indian businessman is a crook—is a worst assumption you can make…
Global Audience & Scale
Millions of users / Billions in revenue
Scale of customer impact served by Anthropic's platform team