Katelyn Lesse and Angela Jiang, who lead Anthropic's platform team, discuss their approach to building Claude's developer ecosystem as an open platform rather than a walled garden.
Platform Vision and North Stars
Anthropic's platform encompasses both external APIs that developers build on and internal infrastructure that powers Anthropic's own products. The team maintains two distinct north stars: internally, they prioritize speed and reliability for shipping AGI-capable products; externally, they aim to give any builder the tools to work with Claude however they want. This commitment to consistency means offering the same primitives to internal and external developers, rather than bifurcating access. The reasoning is practical: since AI capabilities and form factors are rapidly evolving, providing robust foundational tools allows the ecosystem to discover emergent use cases rather than Anthropic locking users into a single vision.
Layered Abstractions
The platform evolves across three abstraction layers. The knowledge layer includes the messages API, system prompts, and standardized components like skills and memory—essentially teaching developers how to use Claude. The execution layer handles the harder infrastructure problems: managing sandboxes, storing sessions for resumable work, and optimizing costs through prompt caching and context management. This layer is accessible through Claude Managed Agents, their main higher-order product. The emerging coordination layer introduces "strategies"—meta-harnesses that assign different jobs to tokens, enabling approaches like having some tokens advise while others execute, or using ensemble-like techniques. The team plans to progressively surface these layers as abstractions evolve with model capabilities.
Ecosystem and Openness
Rather than requiring users to run agents on Anthropic infrastructure, the platform supports self-hosted sandboxes in partnership with Modal, Vercel, Cloudflare, and others. Similarly, MCP (Model Context Protocol) tunnels allow agents to call out to private servers behind firewalls. The philosophy extends to connectors built on MCP specs and Skills, which any developer can use in any product. The team frames this through an electricity analogy: like electrical infrastructure, AI's transformative potential depends on widespread access, standards, and ecosystem partnerships that no single company can achieve alone.
Product Strategy and Experimentation
Anthropic builds first-party products selectively. One approach involves showcasing new form factors—like Claude Design, which explores using code-generation rather than traditional design systems. Another targets specific verticals where agentic automation creates iterative workflows; the team favors "token-heavy" domains where users naturally want to continue after one turn (e.g., coding) over one-shot tasks. They've experimented in finance, legal, and other areas, packaging skills and recipes for developers to use directly or build upon.
Claude Tag, despite appearing as a simple Slack bot, exemplifies this philosophy. The interface is secondary; the real value lies in context engineering and organizational-level harness design—intelligently surfacing the right information, proactively helping, and handling complex workflows. The team envisions this interface becoming fluid: agents appearing wherever humans collaborate (Slack, Teams, email, WhatsApp), with consistent underlying architecture.
Best Practices and Future Focus
For harnesses and context, the team emphasizes prompt caching (critical for cost savings), keeping context windows clean, and evaluations. However, they believe the higher-leverage innovation lies in strategies—composing jobs for tokens rather than further optimizing lower-level execution. As models become more capable, harnesses shift from steering mechanisms to enablers of longer-running, multi-step autonomy.
On token cost optimization, the team advises against capping AI usage entirely. Instead, they encourage designing routers that intelligently match task complexity to model capability within the Claude family, allowing companies to rationalize spending while preserving innovation. Recent customer innovations have included clever context engineering, leveraging computer use for legacy systems without APIs, and exposing agents as MCP servers so multiple agents can orchestrate together.
Looking ahead, the platform team is focused on making composable strategies practical—handling "best of N" sampling and other high-return techniques without requiring custom engineering. They're also addressing enterprise requirements like security, compliance, and modularity, while ensuring weekend developers have accessible, hackable tools. The guiding principle remains consistent: provide primitives and standards so the ecosystem can discover form factors and solutions rather than Anthropic dictating them.