Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden
Training Data
1 DAYS AGO
Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden
Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden
Training Data
1 DAYS AGO
Anthropic's platform leads, Katelyn Lesse and Angela Jiang, discuss their strategy for building a developer platform that supports both external builders and internal product teams. They outline a three-layer stack of knowledge, execution, and coordination, emphasizing that the real leverage lies in the top coordination layer, where meta-harnesses assign different roles to tokens. The conversation covers their philosophy on open ecosystems, the importance of standards like MCP, and their approach to model routing.
The platform's evolution moves from a stateless Messages API to higher-level abstractions that support agentic workflows, including managed agents for enterprises and low-level primitives for AI-native startups. Anthropic favors an open ecosystem, partnering with companies like Modal and Vercel for self-hosted sandboxes, and is not precious about owning the entire stack. They focus on providing sound architecture and standards like MCP to enable interoperable agents. The team discusses best practices for building agent harnesses, including prompt caching and context management, and predicts a shift from token maxing to token rationalization. They advise routing tasks to appropriate models based on complexity and building cost-effective strategies like best-of-n sampling. The goal is to make complex AI agent strategies easy for developers, addressing both enterprise needs for modular, secure solutions and weekend developers seeking open, hackable platforms.
00:00
00:00
Coordination layer uses meta-harnesses to assign token roles
01:49
01:49
Maximum leverage for internal teams to ship AGI-pilled products
02:27
02:27
Supporting builders externally, speed internally
03:54
03:54
Equal primitives for all builders
06:00
06:00
Anthropic now provides out-of-the-box infrastructure for agentic workflows
08:19
08:19
AI-native startups prefer primitives; enterprises prefer packaged offerings.
09:11
09:11
The bottom layer is knowledge, including primitives like the Messages API, tools, skills, and memory.
10:23
10:23
The execution layer involves Claude performing tasks beyond just answering questions.
11:09
11:09
Coordination layer assigns roles to tokens
14:37
14:37
AI is like electricity, a transformative utility.
15:39
15:39
Anthropic favors an open ecosystem over a walled garden
17:12
17:12
Form factors are dynamic and must evolve with AI capabilities
25:02
25:02
The hidden part of the iceberg is the most challenging and valuable.
35:25
35:25
MCP makes agents modular and interoperable
38:13
38:13
Shift from token maxing to token rationalization
