Karan Vaidya

Karan Vaidya is the co-founder and CTO of Composio, alongside co-founder Soham Ganatra. His work centers on connecting AI agents to applications and providing the infrastructure around tool use: authentication, permissions, discovery and execution. He describes Composio as an “agentic tool execution layer,” emphasizing how agents use tools rather than simply how many tools they can access.[2][3][4]

In his March 22, 2026 interview with Nathan Labenz, Vaidya explained two design choices for avoiding context overload. Just-in-time discovery loads tools relevant to the current task instead of exposing the entire catalog. Sandboxes let agents write code to process large workloads, such as batches of emails, while infrastructure primitives handle authentication and supporting utilities. These are his descriptions of the team's architecture; the email example illustrates an intended workflow rather than a measured capacity.[4]

Vaidya also describes a feedback loop around tool execution. When an agent repeatedly fails with a tool or cannot understand it, Composio's internal agent pipeline can generate a revised version and add it to the agent's context. Execution traces supply another kind of improvement: the team converts circuitous paths into reusable skills, containing instructions and scripts for reaching an outcome, and records recurring pitfalls. He distinguishes generic improvements shared across users from personalized upgrades, with skills changing less frequently to preserve repeatable behavior.[4]

For Vaidya, detailed skills also offer a way to reduce dependence on a particular model. He reports creating skills with Opus and then using them with Sonnet for speed, arguing that sufficiently capable models can follow a well-specified path to a similar outcome. He qualifies this view when switching providers: differences in behavior, such as whether a model keeps polling for a result or stops for user input, can prevent a skill from transferring unchanged.[4]

Vaidya emphasizes action-level permissions—for example, allowing an agent to read email without allowing it to send—and hooks around tool calls and returned data that can support guardrails and human review. His operating judgment in March 2026 was that Composio still needed people to supervise agents and make decisions: “we definitely need humans to control the agents.”[4]

His approach to delegation depends on where context belongs. For a small task such as booking an appointment, he favors giving the main agent suitable tools because it already knows the user's calendar and constraints. For exploratory work that consumes substantial context, such as research, he favors subagents that investigate separately and return condensed results. He likewise expects MCP and CLI interfaces to coexist, rather than one becoming the sole way agents access tools.[4]

Vaidya describes personally using an OpenClaw agent for recruiting: finding contributors through open-source repositories, gathering candidate information and conducting outreach on his behalf. On building versus buying agents, he stresses customization and governance. He praises Intercom's Fin for helping companies get started quickly, while arguing that building a support agent gives teams more freedom to define skills, limit tools and connect other applications. Whether that investment is worthwhile, he says, varies by company.[4]

Before this work on agent infrastructure, Vaidya's background included IIT Bombay, Google and Rubrik, which he lists on his personal website. He and Ganatra first met at a Physics Olympiad camp and later became roommates at IIT Bombay. They initially pursued LLM coding agents, then made connecting agents to applications the foundation of Composio's subsequent iteration.[2][3]

In a March 17, 2026 retrospective, Vaidya recalled beginning work with Ganatra in February 2023 and building Composio on nights and weekends alongside full-time employment. He said he left his job in the week his daughter was born, in June 2023. Crediting his wife Mahi's support, he argued that entrepreneurial risk is shared: “Founders talk about risk like it's something they carry alone. But the person next to you carries it too, often more quietly and with less choice in the matter.”[6]

An early pilot made the limitations of the original product concrete. In his November 14, 2025 account, Vaidya said a relationship from his Rubrik days opened the door to an unnamed AI enterprise whose integrations were a bottleneck. The team's early code-generation agent could not help consistently enough. For two to three months, they held weekly calls, wrote and repaired integrations inside the customer's systems, and learned which assumptions needed rebuilding. He credited the customer with helping shape the product and wrote that trust came from an honest acknowledgment: “we didn’t pretend the product was ready.”[5]