Anthropic is building an open ecosystem for AI development rather than locking developers into proprietary infrastructure. In a recent podcast interview, Angela Jiang (Head of Platform) and Katelyn Lesse (Product Lead, Platform) outlined how the company plans to evolve its Claude AI platform beyond simple API access into a coordinated system where different AI models and tools work together seamlessly.

What Is Anthropic's Three-Layer Platform Model?

Anthropic's platform team has designed what they call a "three-layer abstraction cake" to organize how developers build with Claude. Rather than treating the platform as just an interface to access the AI model, the company is deliberately building outward across three distinct layers, each with its own purpose and set of tools.

• Knowledge Layer: Gives Claude the context, tools, and skill definitions it needs to understand a task. This includes the Messages API, tool schemas, Model Context Protocol (MCP) servers, and skills that developers can reuse across applications.
• Execution Layer: Lets Claude run longer and manage complex, multi-step work. This layer handles state management, sandboxes, prompt caching, and context-window management so the AI can complete real work, not just provide information.
• Coordination Layer: Composes multiple AI invocations into orchestrated workflows. This is where developers assign specific "jobs" to different token invocations, such as having one model advise, another execute, and a third verify the results.

Jiang explained the progression: "If you were to look at our road map and project forward a little bit, we'll move more and more from the knowledge layer to the execution layer and from the execution layer to the coordination layer". The company has already moved from knowledge-layer primitives in mid-2025 to execution-layer products like Claude managed agents in early 2026, and is now beginning to expose coordination-layer abstractions.

Jiang

Why Is Anthropic Rejecting a Walled Garden?

Anthropic's leadership explicitly rejected the idea of forcing developers to use only Anthropic's infrastructure. Katelyn Lesse stated: "We aren't precious about you running these things on our infrastructure". This philosophy shapes the company's entire partnership strategy and product design.

Katelyn Lesse

The company has backed up this commitment with concrete partnerships and technical choices. Anthropic has partnered with Modal, Vercel, Cloudflare, and Amazon's new microVMs to offer first-class support for running sandbox execution on third-party infrastructure. The company also built MCP tunnels that allow MCP servers behind a customer's firewall to be called directly from Claude, and created connectors built on the MCP spec so third-party agents can plug into Anthropic's products.

The reasoning behind this openness rests on two core beliefs. First, form factors and technology preferences evolve rapidly. What works today, such as chat-based interfaces, may become obsolete within a year, so locking customers into proprietary infrastructure creates a liability rather than an advantage. Second, the most interesting innovation happens when the broader ecosystem composes its own solutions rather than waiting for a single company to build everything.

How Are Agentic Strategies Evolving?

The most technically ambitious part of Anthropic's platform roadmap involves what the company calls "meta-harnesses" or "strategies." These are design patterns that assign specific roles to different AI invocations within a workflow. Rather than sending one request to Claude and hoping for the best, developers can now compose multiple invocations where each token has a distinct job.

Anthropic's platform team described three generations of agentic infrastructure. Steering harnesses from 2024 to 2025 used heavy scaffolding to force the model along a predefined path, but became largely obsolete as models became more steerable. Execution harnesses from 2025 to 2026 introduced lower-level loops with prompt caching and context-window pruning, and are now commoditized in Claude managed agents. The emerging generation, meta-harnesses and strategies from 2026 onward, will let developers design patterns where a small model executes cheaply, a large model advises, and a reflection loop writes learnings back to memory.

Jiang provided a concrete example: building a bug-hunting agent. Simply sending one agent was not enough, and swapping to a larger model or letting it run longer provided diminishing returns. The real lever was running multiple attempts and taking a best-of-N approach, but implementing that in production is "really, really freaking hard." Anthropic plans to make these strategies reusable primitives so developers can say "I want best-of-N with three workers, and a reflection pass every fifth cycle" without building custom infrastructure.

How to Optimize AI Costs While Maintaining Performance

As enterprises scale their AI usage, cost management becomes critical. Anthropic's platform team described a natural cycle that most organizations experience: token maxing, where teams try to throw the most intelligent model at every problem, followed by token rationalization, where they optimize for cost per unit of outcome.

• Build a Router: Rather than capping AI spend outright, create a router that assesses task complexity and directs requests to the appropriate model. This allows teams to continue getting returns from AI without overspending on expensive models for simple tasks.
• Implement Prompt Caching: Anthropic's platform team emphasized that prompt caching is table-stakes for cost management. Caching repeated context can save significant money without sacrificing performance or quality.
• Manage Context Windows Strategically: Instead of pulling all available data into the context window, clear old tool calls and programmatically retrieve data as needed. This reduces token consumption without limiting the AI's ability to access information.

Jiang stressed an important principle: "What you don't want to do is stop AI usage. If you are getting returns, you are shipping faster, running more operationally efficient, those are gains". The goal is not to reduce AI spending but to optimize how that spending translates into business value.

Jiang

What Products Is Anthropic Launching to Test New Interaction Models?

Anthropic is experimenting with new ways for developers and organizations to interact with Claude beyond traditional API calls. One notable example is Claude Tag, launched in early to mid-2026 as an agent accessible from Slack. The company pushed back against the perception that it is "just a Slack bot." Underneath, it is a heavily engineered context-harness that is proactive and always-on, capable of retrieving information across corporate tools and completing workflows like expense reports without human intervention.

These product launches illustrate Anthropic's broader philosophy: the interface matters less than the underlying engineering. What appears simple on the surface, like a Slack bot, can represent sophisticated orchestration of multiple AI invocations, memory systems, and tool integrations working in concert.

Anthropic's platform strategy reflects a fundamental belief that the future of AI development is not about one company controlling everything, but about building robust abstractions that let the broader ecosystem innovate. By moving from knowledge-layer primitives to coordination-layer strategies, and by deliberately partnering with hyperscalers and infrastructure providers, Anthropic is positioning itself as a platform company rather than a walled garden. This approach may determine whether Claude becomes the foundation for a thriving ecosystem of AI applications or remains confined to a single company's infrastructure.