Before an engineering team gives an AI agent a job, Sunil Mathew wants it to answer a more basic question: does the job need an agent at all?

For the Bengaluru-based founder and lead instructor of The Living Craft, that question is central to responsible design. Some problems are better solved with ordinary software. Where an agent does add value, the next task is to decide what it may do independently and how the system will respond when it gets something wrong.

Sunil brings 26 years of engineering experience, including roles at Google, Amazon, Walmart and startups. Through The Living Craft, he teaches agentic and systems architecture to experienced engineers, architects, engineering managers and engineering leaders. The programme connects building a working system with the reasoning needed to explain and defend its design.

Why engineering depth matters

Sunil describes himself first as an active learner.

He explains, “I am happiest when I am learning something new, and teaching is how I share that.”

Over the last 15 years, he has led teams and hired, developed and assessed senior technical professionals at Walmart, Amazon and Google. In that work, he noticed a pattern: the people who were highly effective  understood complex, large-scale systems in depth. They could explain how a system worked, design its architecture and make decisions about its behaviour.

He believes that depth becomes more valuable as teams build systems in which AI can act. An agentic system uses AI to make decisions and carry out tasks through software tools. Its design must account for the consequences of those actions, alongside the quality of the model’s answers.

That view shapes The Living Craft’s central idea: “AI builds, and the human judges and directs.”

A lesson from a complex membership system

One experience Sunil points to involve automating renewal-related membership decisions at a major technology company.

The system had become increasingly complex as services were added around it. Understanding the existing architecture proved essential to making progress. The team needed to establish which services held particular decisions and why those responsibilities had been placed there.

Sunil says that understanding those earlier design choices helped the team save time and cost.

He sees a similar challenge in agentic systems. The model attracts attention, but engineers must also understand what an agent can reach, where decisions are made and what happens after an incorrect action. Those relationships determine how the wider system behaves.

For Sunil, the example explains why learning architecture requires examining the decisions already present in a system..

Asking about the failure path

Once a team establishes that an agent is appropriate, Sunil turns to its authority.

His next question is “what is the agent allowed to do without a human, and what happens the first time it gets that wrong?”

The question connects permissions to failure handling and review. A design needs to show which actions require a person’s approval, how a mistake will be detected and who is responsible for the response. It also needs to account for the cost of being wrong.

These questions bring the discussion back to behaviour that an engineering team can examine. A successful demonstration shows one path through the system. Reviewing the failure path asks what happens when the assumptions behind that demonstration no longer hold.

Sunil puts it plainly: “The failure path is where the real engineering is.”

Learning with a working system on screen

The Living Craft puts those questions into practical work. The programme includes 30 live hours with Sunil, plus independent work, with weekly live sessions centred on a working system.

Members work on their own agent system and write decision records that can be read in an architecture review. They test their systems to find failures, trace the cost of individual calls and change the model to examine what breaks. Paired design reviews give them practice explaining their choices and examining someone else’s reasoning.

The system provides something concrete to return to during discussion. Evaluation and feedback can reveal a gap in the design, which members then investigate and revise. Continuous feedback connects the system they build to the judgment they are developing.

Sunil wants members to carry a particular habit back to their teams. Before shipping an agent, they should be able to explain what it is allowed to do, how they will know when it has gone wrong and what it costs to operate. They should also be able to defend those answers in a design review.

Bringing the same questions into enterprise work

Alongside The Living Craft, Sunil advises Indian enterprises, particularly those in regulated sectors, as a Fractional Chief AI Officer. He also runs a fixed-scope AI Readiness Assessment.

His teaching and advisory work share an emphasis on deciding where AI is useful and understanding the responsibility that follows its use. For an engineer, that can begin with a single design decision. For an organisation, it includes how those decisions are reviewed and directed.

The Living Craft gives experienced technical professionals a place to practise that judgment through building, examination and revision. Its starting question remains a useful one to take into the next design review: does this problem need an agent, and what responsibility will the team take on if it does?

Programme information is available at The Living Craft.