← All speakers

Sagar Batchu

Conference affiliation: CEO · Speakeasy · 2026

Sagar Batchu is co-founder and CEO of Speakeasy, which builds API development tools and infrastructure for governing AI agents’ access to enterprise systems. His work spans two connected problems: helping developers call APIs through maintained software development kits, and giving companies control over which tools agents can use and how their actions are investigated.

On this page

From physics to developer infrastructure

During his final year at university, Sagar decided against continuing in physics research. He took more computer science courses and became interested in building infrastructure that other developers could build upon.

At LiveRamp, he worked on internal developer tooling and data infrastructure, moved to the UK, and helped establish an engineering team there. Building a new operation inside an established company gave him experience developing prototypes into products for customers. In his conversation with APIs You Won’t Hate, he connects that experience to an ambition for his own company: making infrastructure available inside large engineering organizations accessible to developers elsewhere.

Sagar co-founded Speakeasy with Simon Yu. The company addressed the work between discovering an API and making a successful request: developers switching between documentation, an editor and support conversations to get an integration running. Managed SDKs turn an API into code developers can use in their own language, while relieving the API producer of maintaining separate client libraries.

In October 2024, Speakeasy announced a $15 million Series A, led by FPV Ventures with participation from GV and Quiet Capital. Sagar described a remit extending beyond generating an initial library: keeping documentation current, catching accidental breaking changes and preventing SDKs from falling behind the APIs they expose.

Choosing a shared standard

Speakeasy considered creating its own domain-specific language for configuring SDK generation. Sagar and the team chose OpenAPI instead, accepting its verbosity and complexity to work with existing frameworks and tools. His case for that decision preserves the tradeoff: vendor extensions fill gaps around pagination, streaming and retries, but introduce some dependence on Speakeasy’s conventions. He favors extending a shared standard over asking customers to adopt another specification.

In September 2026, Sagar announced a collaboration with Google to open-source Speakeasy’s client generator. The teams worked together on SDK migration, including preserving error hierarchies and streaming behavior—details that determine whether existing applications can move to a new client without changing how they handle failures or consume responses. The release made SDK, CLI and MCP server generation available in the open.

From building agent interfaces to governing their use

By early 2025, Sagar was describing agents as a new kind of API consumer. A flight-booking assistant needs precise descriptions of inputs and outputs to execute a sequence of calls; a documentation website designed for a human does not supply everything the model needs. His account of agentic API interactions extends Speakeasy’s integration work to interfaces that models can interpret and use.

Customer experience subsequently changed his view of where companies needed help. In March 2026, Sagar revised Speakeasy’s original MCP thesis. The team had expected building Model Context Protocol servers to be the difficult part. Customers could often produce a server readily, yet deployment stalled when they needed to get it approved, understand its failures or distribute it to employees. Speakeasy shifted its emphasis toward managing servers, including those built elsewhere.

The resulting Speakeasy MCP Gateway puts authentication, permissions scoped to individual tools, telemetry and audit information between agents and servers. It addresses both access and investigation: which tools a person’s agent may invoke, and what an operator can examine after an invocation fails or causes harm.

Sagar makes the stakes concrete in A Valid Key Let an AI Agent Delete Production, using a coding agent’s accidental deletion of a production database to explain the limits of a valid credential. His proposed controls associate a request with its owner, inspect the resource and intended action, enforce policy before execution, and retain the decision for later investigation. The credential permits a connection; the intended action still needs its own authorization.

1 conference talk

Loading more talks and highlights…

References