Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden

Summary

Katelyn Lesse and Angela Jiang describe Anthropic’s developer platform as the infrastructure layer behind both Claude-facing external APIs and Anthropic’s own internal products. Their architecture roadmap has three layers: knowledge, execution, and coordination, with “strategies” envisioned as a meta-harness that allocates different jobs to different tokens. They trace the platform’s evolution from a mostly stateless Messages API toward agentic systems with long-running context, sandboxes, session resumption, prompt caching, and managed infrastructure. On ecosystem strategy, they argue Anthropic should stay open via standards like MCP, Skills, connectors, and self-hosted sandboxes, but they do not believe harnesses should be model-agnostic; instead, they should be tuned to a specific model family. The broader thesis is that the highest leverage now comes from orchestration, context engineering, and cost-aware routing, not just bigger models or more prompt steering.

Notes

• Anthropic’s platform serves both external developers and internal product infrastructure.
• Internally, the north star is leverage and speed for product teams.
• Externally, the north star is to give any builder the tools to work with Claude.
• The team tries to keep internal and external primitives aligned.
• Their rationale is that model capabilities are changing too fast for durable form factors.
• They dogfood internally while opening early access externally to widen feedback.
• The platform originally looked like a stateless Messages API.
• Customers pushed the company toward longer-running and remote agent contexts.
• Agent infrastructure now includes sandbox governance, security, and resumable transcripts.
• Prompt caching, context-window management, and cost control are explicit harness concerns.
• The execution layer is low-level harness plus managed infrastructure.
• Managed Agents is the current product name for that execution layer.
• The coordination layer is where Anthropic is starting to think about strategies.
• Strategies are framed as a meta-harness above execution.
• Anthropic has standardized knowledge-layer primitives over the past year-plus.
• The company wants products and agents to plug into each other across vendors.
• Skills and MCP are positioned as core ecosystem standards.
• Anthropic has built self-hosted sandboxes as a first-class integration path.
• MCP tunnels let customers reach servers behind firewalls.
• They treat ecosystem openness as more important than owning every layer.
• Cloud Tag is mainly about the architecture underneath the interface.
• Anthropic’s aim is for an always-on company agent that feels like a coworker.
• Prompt caching and context hygiene are baseline best practices.
• The guests think steering scaffolds matter less as models get more steerable.
• High-verification domains need careful error-handling between model steps.
• Anthropic gives builders many knobs today, but wants outcome-and-budget control eventually.
• Innovation is moving into distributed context collection and permissioning.
• Healthcare is a compelling use case because many systems still lack APIs.
• Customers are exposing agents through MCP so other agents can call tools on them.
• Manufacturing is emerging alongside coding as an AI adoption category.
• They describe the market as moving from token maxing to token rationalization.
• The next optimization dimensions after intelligence are cost and speed.
• Anthropic does not want customers to stop AI usage in response to spend.
• Instead, they want task-based routing across model sizes and costs.
• Anthropic’s harnesses should be tuned to the model family they use.
• The company is not pursuing arbitrary cross-model routing as the default abstraction.
• There is still room for smarter model mixing within a family.
• Best-of-N is treated as a third lever beyond bigger models and longer runtime.
• Anthropic is pushing into coordination-layer tools for composing strategies.
• They want to maximize intelligence per dollar while preserving enterprise readiness.