Building APIs for Developers and AI Agents

I love to say APIs are the dendrites of the internet. Servers fire by calling each other, and that synapse is almost always an API. Every company is becoming an API company, and each needs SDKs, docs, versioning, and more so teams can focus on core capabilities and clean interfaces, not low-level details.

If you have been building with APIs, you know they are the connective tissue of modern software. As they multiply, so do challenges around SDK quality, documentation, versioning, and now interfaces for AI agents. In this episode, Stainless founder Alex Rattray joins a16z partner Jennifer Li to discuss Stripe-level developer platforms for everyone, why SDKs are often the API, how MCP reframes APIs for large language models, and designing for agents as first-class users. Let us get into it.

At Stripe I helped redesign API docs and realized the SDK is the API for many developers. I built a code generator from OpenAPI to great Python and TypeScript libraries, then saw most companies could not maintain that quality. That is why I started Stainless: bring a Stripe-grade developer platform in a box to every API company.

How did you approach SDK generation given Stripe’s quality bar?

Robust first: resilient networking, clear errors, telemetry, scale. Polished next: idiomatic per language with great inline docs and types. Python evolved a ton—types, async, pattern matching—so we chased what feels native to a Pythonista. With MCP, the new question is what feels idiomatic to an LLM using a large API. That is a fresh, fun research problem.

You call APIs the dendrites of the internet. What is MCP to you, and how does it change the landscape?

APIs are the synapses of the internet. MCP is the right interface for the GPU brain—a new sense for software. We had eyes and fingers via dashboards and automation; MCP feels like growing another limb so agents can perceive and act across existing capabilities.

Designing for large language models as users, what changes for your product?

I treat MCP as the SDK between an API and an LLM. The hardest problems are context and tool overload. Naively exposing every endpoint and parameter from OpenAPI blows the context window and overwhelms agents. You need progressive disclosure of endpoints and parameters with minimal round trips.

Do you have solutions for context limits and tool explosion?

Three approaches today. One, filter tools up front by resource or read versus write via CLI flags or URL params. Two, dynamic mode with three O of one tools—list operations, describe an operation, execute—trading a couple more turns for huge context savings. Three, JQ-style filters on responses so agents can request only the fields they need, keeping payloads small and usable.

Some think MCP reduces the need for great SDKs and docs. I think the opposite. Your take?

I agree. Agents tend to install SDKs, but they often grab old versions and hallucinate APIs. With MCP we can feed authoritative, versioned SDK reference docs directly to the model so it uses the right surface. Strongly typed SDKs then give type checking and tight feedback loops, so agents iterate safely without risky one-shot HTTP calls. That means faster integrations and more confidence.

Look five to ten years out. As agents become common, what does the agent experience look like?

Code gets more declarative and dry, and backends less spaghetti. Clearer, prescriptive API design standards emerge so LLMs can follow company rules and frameworks, declare endpoints and parameters, write the business logic, and let humans review quickly.

And the split of work between humans, agents, and platform?

Humans specify high-level intent. AIs produce the mid-layer business logic. Platform code should be stable frameworks and libraries. Stainless aims to handle the low-level infrastructure around the API so teams focus on what matters.

Any learnings from working with customers like OpenAI, Anthropic, and Cloudflare?

Fundamentals endure: be thoughtful, design for the user, expose capabilities cleanly, and remove low-level headaches. Whether it is LLMs, cloud, or money movement, teams want to focus on core strengths and present crisp, reliable interfaces.

Thank you so much, Alex. This was wonderful.

Thank you, Jennifer. This was fun.

As APIs become core infrastructure for both developers and AI agents, the quality of SDKs, docs, and interfaces will only grow in importance. If you enjoyed this conversation, please rate and review the podcast and share it. Stay tuned for more on the future of software, infrastructure, and AI. This content is for informational purposes only and not legal, business, tax, or investment advice, nor for evaluating any investment or security. See a16z dot com slash disclosures.