Ask a room of developers about agents today and you will get plenty of answers. Almost everyone has built one, and most have a demo they are proud of.
The harder question is what happens after that, when the agent has to hold up against real customers, real data, and real consequences. At DevSparks Chennai 2026, Arvind Purushothaman, Global Cloud Consulting, Microsoft Practice and Alliances Leader at EY GDS, opened with a version of it. He asked how many people in the room had built an AI agent, and plenty of hands went up. Then he asked how many had built one that made it to production for a business-critical system. "I think that's where the numbers dramatically drop," he said. "Not because of anybody's fault, but that's how it is."
That gap is the whole problem, and it is not a model problem. Cloud, low code tools, and open model APIs have pushed the entry barrier close to zero. What has not changed is the difficulty of industrializing what you build. "Access to intelligence is not the bottleneck, but the ability to industrialize is going to be the bottleneck," said Arvind Purushothaman.
From the nets to the match
Drawing on a cricket analogy, he said that hackathon builds are like net practice, where a few lines of Python and an API key let you rehearse your reverse scoop all afternoon. The match is different. You need 12 runs off six deliveries under pressure, and the strokes only matter alongside the ability to read the situation.
For agents, that pressure arrives when the demo meets regulated data, compliance requirements, integration points, and business-critical workflows. Skip the rigor and the exposure is concrete: a bad transaction, financial loss, reputational damage, security and privacy failures. The question itself changes, Purushothaman said, from does it work in a demo to "can we trust it repeatedly under real operating conditions."
More autonomy means more engineering
Traditional software ran on deterministic rules and stage gates that everything had to clear before production. Agents behave differently. They interpret context, reason about a goal, and decide for themselves which tools to use to reach an outcome.
The instinct is to read it as less work for engineers. Purushothaman argued the opposite. More freedom for the agent means more engineering responsibility for the people around it, because the system has to constrain, restrict, and verify what the agent does. "Autonomy and controls … they are not the opposite of each other in this world," he said.
He offered three maturity levels to locate yourself against. Task-based agents automate specific steps and return an output. Goal-oriented agents are handed an objective, such as how much credit limit a customer should get, and assemble their own data and policies to reach a recommendation. The third horizon is humans and agents as one team, with judgment reserved for critical calls.
Ontologies, controls, and platform
Three foundations make production deployment possible, and they only work together. Domain context and ontologies codify business knowledge, taxonomy, and decision logic, and mature vendors now ship them by industry. Controls evaluate outcomes for quality, bias, and risk, and produce traceability. The platform makes execution repeatable and scalable. Ontologies without controls create chaos, controls without a platform are not repeatable, and a platform without domain context is generic.
The same logic scales up.
The model is only one part of the system, so the engineering has to cover data context, cloud and compute, integration APIs, security and identity, responsible AI, and FinOps discipline around cost. "The outcome is only as strong as the weakest link in your chain," Purushothaman said.
Every agent needs a passport
Take a customer asking a bank to raise their credit limit. The agent has to establish identity, retrieve the applicable risk policy before anything else so it knows which boundary it is operating under, pull the relevant data, and interpret context. The recommendation has to stay inside policy guidelines, and if the risk is high, it should transfer to a human. Every step has to be auditable. Without that preserved history, he said, it is not a production-ready system.
Hence the agent passport. Every agent gets an identity, a named business owner and technical owner, defined permissions covering where it can operate, and escalation points. Organizations already carry deprecated software they cannot account for, and agents will multiply that problem faster. Production systems also need circuit breakers: anomaly detection, the ability to halt execution, roll back, and escalate to humans.
What value actually looks like
Asked how to measure return, Purushothaman pushed back on counting agents deployed and tokens consumed. The number that matters is cost per governed business outcome, counting cloud and agent running costs, the human effort still in the loop, and the risk profile. "You don't want to over-engineer a simple ask, and you don't want to under-engineer a complex ask. It should be right for the outcome."
Testing, he added, has to shift from fixed test cases toward simulating production like runtime conditions. Asked whether robustness holds as attackers keep evolving, he conceded that the how will change while the fundamentals will not. "We talked about the why. We talked about the what. The how will evolve."
He closed on the sharpest line of the session. "Intelligence is abundant, but trust, accountability, and repeatability, we must still engineer."