Is "Agentic in Production" a product or a framework?

Neither. It is a standard, a set of four tests an agent has to pass before it does real work, and a body of writing about how to pass them. Ailoitte uses it on client work; anyone can use it.

Does it apply to simple automations, or only to autonomous agents?

It applies to anything where a model decides what to do next. A single prompt that drafts an email does not need it. A workflow where the model picks tools, retries and acts on systems of record does, whatever you call it.

What is the difference between a pilot and production?

A pilot has permission to fail quietly. Production has someone who is woken up when it fails. If nobody would be paged, it is still a pilot.

Agentic in Production — Sunil Kumar

Most enterprise agentic AI dies between the demo and production. The demo works. The pilot mostly works. Then someone asks who is accountable when the agent is wrong, what it costs at ten thousand runs a day, and what the auditors will see, and the project quietly stops. Agentic in Production is my name for the work that gets an agent across that gap, and for the standard it has to meet on the other side.

What the term rules in and out

An agent is in production when four things are true at the same time.

Its failures are bounded. There is a written list of actions the agent may take and a written list it must refuse, and the second list is enforced in code, not in the prompt.

Its behaviour is observable. Every run leaves a trace that a person can read after the fact: what came in, what the model decided, which tools it called with which arguments, what came back, what went out.

It is owned. One named person is accountable for it, can switch it off in under a minute, and is the one who gets paged.

It has a budget. Someone has decided what a single run may cost, what the month may cost, and what the work is worth. The agent stops when the numbers disagree.

A demo has none of these. A pilot usually has one or two. Production needs all four, and the last two are organisational work, not model work. That is why so many technically sound agents never ship: the engineering team can do the first two alone, and nobody owns the other two.

Production-grade agent: An agent whose failures are bounded by an enforced refusal list, whose every run is traceable, which has a named owner with a kill switch, and which runs inside a budget it cannot exceed.

Why the model is rarely the problem

When an agent project stalls, the post-mortem usually blames the model: it hallucinated, it was too slow, it was too expensive. In my experience the model is the easiest part to swap. What actually failed was one of the four tests. Nobody had written the refusal list, so the first embarrassing output ended the pilot. Nobody could reconstruct a run, so the compliance team said no. Nobody owned it, so it drifted. Nobody budgeted it, so the invoice for month two was the last one anyone approved.

Fixing the model does not fix any of those. Passing the four tests fixes all of them, and it makes the model choice a routine decision you can revisit every quarter.

Where the phrase comes from

I have spent the years since 2017 building Ailoitte into a studio that ships software people rely on: more than 300 products, more than 50 million users, in healthcare, fintech, enterprise and government. The agent projects that succeed look like the software projects that succeed. They are boring in the right places. They have a runbook. Someone is on call. Agentic in Production is what that looks like when the software includes a model that decides what to do next.

The standard is not theoretical. The 306 agents in Leverge, Ailoitte’s platform, pass the four tests by construction: a refusal list per agent, a trace for every run, a named approver for any irreversible action, and a budget per step. What running them taught me is its own essay.

Where to start

The essay on why most enterprise AI agents never leave the demo walks through the three gaps a project dies in.
The essay on what “production-grade” actually means turns the four tests into things you can check.
The reference architecture playbook shows how to build an agent that passes them.
How we deliver it at Ailoitte: AI Velocity Pods. How the lifecycle changes: AI-DLC.

For background on the field itself, the Wikipedia entry on intelligent agents covers the classical definition, and Anthropic’s note on building effective agents is the clearest short description of the current engineering patterns.