The Harness Problem: Cleartax's CTO on What It Takes to Ship an Agent

This piece draws from a fireside chat and AMA session, "Failures, Surprises, Advice," at AISRENext Bengaluru. Someone in the audience pushed the panel on a question that comes up whenever teams evaluate AI agents: with tools like Claude available, why not just build the agent yourselves instead of buying one?

Suvesh Malhotra, CTO at Cleartax, has heard the same question from his own customers.

Cleartax Gets Asked the Same Build-vs-Buy Question It's Answering Here

Cleartax isn't building agents only for internal use. The company builds a CFO office platform, and its own customers regularly ask whether they could build the same agents themselves.

"You can definitely build agents, but agents have to work with a high amount of accuracy. It has to work correctly. It has to work all the time. So you have to build a whole harness around it for it to actually work in the way you want, for the outcome to be driven."

Suvesh isn't describing a thin layer of prompt engineering wrapped around a model. A harness, in his framing, covers everything that has to exist around an agent so that "it usually works" becomes "it works every time, on every customer's data, without someone checking behind it."

The Harness Problem Behind "It Usually Works"

"In the past also, people could have built internal software, and a lot of companies used to build. But that's the difference, actually getting the software, and now agents, from somebody who is an expert has a lot of value."

Companies have weighed building internal tools against buying them for decades. What's different with agents is the size of the gap between a working demo and something that holds up in production, harder to close without specialized expertise than it was for most traditional software.

Why Some Agents Still Get Built In-House at Cleartax

A company arguing agents are hard to get right, and yet Cleartax still builds some of its own. Suvesh pointed to two reasons.

Cleartax launched its first agent, a tax assistant on WhatsApp, two years ago, before most agent platforms and tooling existed in mature form. Building in-house at that point wasn't a strategic preference so much as the only option on the table. Beyond timing, there's a second reason that's held even as the market matured: as a startup, Cleartax keeps some agent-building in-house not to cover every use case, but to learn the mechanics of building agents before it ships that expertise to customers.

"So those are the two reasons why we are building it. But at the end of the day, I think there's a lot of space for people who build it for a specific reason, because they can do it better than others."

Where Building Ends and Buying Begins

Building an agent to learn how agents work is a different call than building one because a team assumes it can match a specialist's accuracy and reliability at scale. Cleartax treats the first as a deliberate investment and the second with real caution, a more specific position than a blanket "always build" or "always buy."

At StackGen, Aiden is built by a team whose job is agent reliability and governance at scale, the same harness problem Suvesh describes, so customers don't have to build and maintain that discipline themselves. Try Aiden for SRE Community Edition for free today.