Karan Vaidya uses an AI agent to recruit. Not in the loose way people mean when they say their calendar app has AI features — in the literal sense: his OpenClaw instance reviews GitHub repositories, identifies strong commits, surfaces the best contributors, enriches their profiles with location data and email addresses, and sends outreach on his behalf. "It has emailed and I've gotten like in last week or two week, I've gotten, let's say like 30, 40 calls set up," he said on the Cognitive Revolution podcast. "That whole thing is fully done by my agent and Composio," he added (source).
That is intent-based delegation. Not "here are the steps to find a good engineer." Just: here is the outcome I want.
The distance between intent and command is where agent autonomy lives — and where it gets interesting in ways that are not always flattering. When you tell an agent what you want rather than how to do it, you are trusting it to fill in the gaps. Most of the time that works. Sometimes it fills in the gaps wrong, and by the time you find out, the agent has already emailed 40 people.
Vaidya frames this as a phase transition — people are becoming more intentful with these agents, he said, and tend to give the outcome they want rather than the steps to get there, letting the agent figure out the path on its own. His own hiring agent is the example: it decides which repositories to search, how to rank contributors, what constitutes a useful signal — and acts on all of that without a human in the loop for every decision. The human reviews the calls. The agent decides who gets emailed (source).
The practical question is what happens when intent and command diverge in ways the human didn't anticipate. Vaidya described a world where agent profiles can be tuned with granular access controls — one agent might get read-only access to your entire email history so it can do research, but cannot send anything; another might have send permissions but limited data access. That compartmentalization is the proposed answer to the divergence problem: define the intent envelope tightly enough that the agent can't wander far.
But envelope design is a hard problem. The failure mode isn't usually that the agent ignores you — it's that it obeys the wrong version of what you meant. Vaidya's hiring agent apparently doesn't email people by mistake. But the architecture that prevented that had to be built, and not every team building agent workflows is building it consciously.
The divergence between what you said and what the agent understood is also where skills — Vaidya's preferred unit of reusable agent behavior — either help or hurt. Skills encode trajectories: the path an agent should take to reach an outcome. When a skill is well-written, the agent follows the intended path even if the model underneath changes. Vaidya claims skills built with one frontier model work 90 to 95 percent of the time when swapped to a cheaper model. The 5 to 10 percent gap is where behavioral quirks baked into the skill by the original model create drift. Vaidya attributes these gaps to nuances baked into the skill by the original model — reflecting how that model operates and not being fully portable to a different model. The skill encodes not just the procedure but the procedural style — and that style is not fully portable (source).
This is the more honest version of the "skills as lock-in defense" argument. Vaidya's position is that skills reduce provider lock-in because good instructions work across models. The counterposition is that skills encode model-specific behavior in ways that are not always visible, and that 5 to 10 percent gap is the surface area for intent drift. The same dynamic that makes skills portable makes them brittle in ways that are hard to see until the agent is already doing something you didn't intend.
The Intercom Fin question is the clearest example of where this gets commercial. Fin is a product Vaidya explicitly calls "great" — Intercom built an AI support agent and sells it as a flat-rate per-ticket product. Composio has 133 tools for Intercom. A company that wants to customize their support agent — different escalation logic, tighter integration with internal systems, different cost structure — can in principle build it themselves using those tools. "The customizability is what people prefer when making that build versus buy decision," Vaidya said. He expects that as models improve, more companies will inch toward build (source).
The economics are not subtle. Fin's $0.99-per-ticket pricing — confirmed on Intercom's own pricing page and help documentation — against the token cost of running the same workload through Composio's tools is a gap worth examining for any company at scale. Vaidya's own estimate for his three-person engineering team's token bill — $100,000 per month for the agent pipeline — suggests that at scale, token costs can be substantial. The customizability argument is real: owning the skill means you control escalation logic, integration depth, and behavior under edge cases in ways a flat-rate product doesn't offer. Whether the economics pencil out depends on ticket volume and engineering capacity — and Vaidya is clear-eyed that for many teams, the build path is not worth it.
What Vaidya is describing, across all of these examples, is not automation in the traditional sense — the mechanical execution of pre-specified steps. It is the delegation of problem-solving to a system that interprets intent and decides how to act. The hiring agent isn't following a script. Fin isn't a decision tree. The skill isn't a macro.
This is a different kind of knowledge work. The human sets the outcome, reviews the output, and handles the exceptions. The agent handles the path. Whether that path stays close enough to what the human actually wanted depends on how well the intent envelope was defined — and on how honestly the system acknowledges the gap between what was said and what was done.
The 30 to 40 calls set up from his agent's outreach are the success case. The failure case is out there somewhere, running on the same architecture, following the same logic, equally autonomously. The question for the people building and deploying these systems is whether they know which one they have running.
Sources: Cognitive Revolution podcast, episode with Karan Vaidya, CTO of Composio, March 22, 2026 — Cognitive Revolution
Correction: An earlier version of this article mischaracterized the mechanism behind Vaidya's hiring agent's results. The transcript shows Vaidya describing his agent emailing candidates and subsequently getting roughly 30 to 40 calls set up — not receiving inbound calls unprompted, and not the agent directly scheduling outbound calls. This version corrects that paraphrase. Additionally, two direct quotes in this article — a passage about people becoming intentful with agents, and a passage about skill nuances — have been rewritten as paraphrase. They were presented as direct quotes but did not match the transcript exactly; exact transcript quotes are used where available. AWS, Zoom, and Airtable as customers — cited in Vaidya's own statements on the podcast — are unverified by independent sources and have been removed from this version. Glean was independently confirmed as a Composio customer; the other named customers were not.
