Fireside Chat, Interview
Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden
- Abstraction Roadmap: Anthropic is shifting its platform focus from the knowledge layer to the execution layer, and eventually to a coordination layer.
- Three-Layer Model: The platform architecture is defined by:
- Knowledge Layer: Primitives regarding model capabilities, including specific API shapes, tool calling formats, skills, and memory standards (e.g., MCP).
- Execution Layer: Infrastructure for running work beyond simple Q&A, including "Cloud Managed Agents," sandboxed environments, session storage for resuming long-running tasks, and prompt caching.
- Coordination Layer: "Strategies" or meta-harnesses that assign specialized roles to tokens (e.g., advising, executing, dreaming) to compose orchestrated workflows.
- North Star Goals: The team operates with dual objectives:
- Internally: Providing maximum leverage and speed for shipping "AGI-pilled" products.
- Externally: Democratizing access to tools so any builder can create custom software, effectively making the last mile of custom development economically viable.
- Product Philosophy: Anthropic maintains a unified platform for both internal and external users to avoid "bifurcation," ensuring consistent primitives and preventing over-optimization for specific internal requirements that limit external generalizability.
- Form Factor Evolution: The team treats AI form factors as dynamic rather than static; they frequently experiment with internal projects (e.g., "Labs") to test new ways of interacting with models, even if those projects are never shipped externally.
- Vertical Strategy: Product development prioritizes "token-heavy" verticals where AI usage is iterative and generates compounding value (e.g., coding) rather than single-turn tasks, with specific focus on finance and legal due to high stakes requiring specialized verification logic.
- Claude Tag Launch: This product demonstrates an "org-level harness" where the interface is secondary to the underlying context engineering and proactivity, allowing non-technical users to delegate complex workflows via simple Slack interactions.
- Token Rationalization Trends: With "token maxing" occurring, the focus is shifting from raw intelligence to cost and speed optimization via model routing.
- Model Routing: Anthropic encourages customers to route tasks to different models within the Claude family (e.g., Opus vs. Haiku) based on complexity rather than mixing providers, believing harnesses should be tuned to specific model families.
- Ecosystem Interoperability: The platform supports a non-walled-garden approach by:
- Offering self-hosted sandbox options via partners like Modal, Vercel, Cloudflare, and AWS.
- Supporting MCP tunnels to connect to servers behind firewalls.
- Encouraging third-party agents to expose MCP servers, enabling cross-agent tool calling and modularity.
- Innovation Observations: The most significant customer innovation currently occurs in the "context and connectivity layer," including connecting legacy systems without APIs via computer-use agents and building custom software development lifecycles (SDLC) on top of the platform.
- Future Development: The roadmap prioritizes making "strategies" and "jobs" accessible to the ecosystem to help users maximize intelligence per dollar and optimize token budget allocation without building custom infrastructure.
- Developer Experience Focus: Efforts are being made to solve enterprise "table stakes" issues (security, compliance, integrations) to unlock the platform's advanced capabilities for a broader range of user personas, from weekend developers to large organizations.