the living craft
00 An application-only program in agentic & systems architecture

Build systems
that outlive
their makers.

I'm Sunil Mathew. I have spent 26 years building and leading engineering at Google, Amazon, and Walmart. Here I teach the craft of agentic and systems architecture, live, to a small cohort.

Apply for the October 2026 cohort  Ask about the cohort

Inaugural cohort · starts October 2026 · 8 seats, capped · applications reviewed personally

Sunil Mathew
Sunil MathewEngineering leader · Google, Amazon, Walmart
26 years, shipped at
01 Who this is for

For engineers who've hit the ceiling of what tutorials teach.

You can build. What you want now is judgment. That comes from shipping hard things and living with the consequences. It is not in a course catalogue. It transfers from one person to another.

// 01

Tech Leads & Staff Engineers

You make the architectural calls your team builds on. You want them to still look right in two years.

// 02

Senior Engineering Managers & Architects

You set technical direction for agentic and distributed systems, where the failure modes are subtle and expensive.

// 03

Senior Engineering Leaders & Directors

You own outcomes across teams. Titles differ across companies. What you need is the depth to back the decisions that matter.

02 An illustrative example

Follow a decision beyond the demo

Imagine a returns assistant that reads a policy and recommends a refund. Now give it a tool that can issue the refund. The system can move money, and its design needs to account for that responsibility.

What evidence permits the action? How does it recognise a repeated request? What happens when the policy is unclear or a tool result is uncertain?

This illustrative example shows the kind of connection you will practise: from a design choice to behaviour you can examine.

  1. Reads The policy Gathers the evidence it will rely on.
  2. Recommends A refund A person can still read this and decide.
  3. The boundary What permits the action? Evidence, limits, repeated requests, an uncertain result.
  4. Acts Issues the refund The system moves money. Its own controls, not the prompt’s.

03 A second illustrative example

See what a review can change

Imagine a checker rejects an agent's result, and the workflow repeats the same attempt. A useful review asks what new evidence the next attempt has. If nothing changes, the next design question is how to recover or stop.

This is an illustrative teaching example. It shows how a question can lead to a change you can build and examine. The aim is to develop how you reason, alongside what you produce.

04 What you develop

Develop practical design judgment

Build a working agentic system and connect its behaviour to the architecture behind it. Practise explaining why a boundary exists, what evidence supports a decision and what you would change next.

Develop how you design and critique agentic systems, review a team’s proposal and provide useful guidance.

05 The transformation

What you'll be able to do.

By the end, the work shows up in how you decide. Under real constraints, with real stakes.

  1. A.

    Design agentic systems that hold up in production

    Not demos. Architectures with bounded failure, observability, and a cost you can defend.

  2. B.

    Prove your system works, instead of hoping it does

    Build the evaluation harnesses, test sets, and quality gates that turn "it works" into a measurement. Catch regressions before they ship.

  3. C.

    Engineer for a model that won't behave the same way twice

    Guardrails, fallbacks, and a clear view of how much damage one failure can do. This is reliability for systems that are probabilistic by nature.

  4. D.

    Break your own system before someone else does

    Threat-model it, red-team it for prompt injection and data exfiltration, and hold the line between instruction and data.

  5. E.

    Govern an AI-native team

    Risk-tiered review depth, accountability for the code the AI wrote, and the judgment for where the human directs and where the human steps back.

  6. F.

    Make the irreversible calls with confidence

    Read trade-offs, scale, and consistency the way a craftsperson reads the grain of wood. You know where it splits before you cut.

Then carry it forward. The real measure of having learned the craft is being able to hand it on.

06 The work

What the work explores

  1. 01

    Purpose and constraints

    Frame the task, the context and the choices the system needs to support.

  2. 02

    Tools and authority

    Examine what the system may do, the evidence it needs and where a person should decide.

  3. 03

    Evaluation

    Inspect behaviour and failures, and connect findings to a useful design change.

  4. 04

    Reliability and cost

    Consider uncertain results, retries, review effort and the trade-offs around the model.

  5. 05

    Design review

    Explain the architecture, listen to challenges and identify the next useful piece of work.

07 Inside the program

A small set of deep things, done properly.

Six weeks. Live, in a small cohort. Built around your real systems, not toy problems.

M1Week 1

Foundations of durable architecture

How agentic systems actually fail, and how to design so that the failure is bounded, visible, and boring. These are the mental models the rest of the program builds on.

M2Weeks 2–4

Agentic systems you'd put your name on

Multi-agent orchestration, RAG, and tool boundaries. Then the part most courses skip. Evaluation harnesses that prove it works. Reliability engineering for systems that do not behave the same way twice. And attacking your own system, to test it for prompt injection and data theft. Demos are easy. Systems you would run in production are not.

M3Week 5

Scale, consistency & the irreversible trade-offs

Reading the CAP trade-off in real systems. Capacity and cost under load. And the architectural decisions you cannot take back, made with the judgment to know which way they will break.

M4Week 6

Your system, reviewed in the room

You bring a real architecture. We pressure-test it together as a cohort. We cover the design, the failure modes, the evaluation strategy, and the governance around it. That governance means review depth matched to risk, and accountability for the code the AI wrote. This is a senior review the way it should feel.

08 How it runs

Build, examine and revise

The working system gives the discussions something concrete to return to. Build an approach, examine what happens and use continuous feedback to understand strengths, gaps and follow-up work.

Industry-specific practice is the default. Company-specific work needs a separate conversation about scope and permission.

09 Who teaches it

Work with Sunil Mathew

Build
Engineering systems
Lead
Guide technical decisions
Review
Examine evidence and risk
Teach
Transfer design judgment
Sunil Mathew
Sunil Mathew · Bengaluru, India

Sunil brings engineering and leadership experience from Google, Amazon, Walmart and startups. His focus is the reasoning behind a system’s design and the evidence that helps a team make its next decision. His teaching connects a concrete example, a question about the design and a change that can be examined.

He has spent twenty-six years building and leading engineering at Google, Amazon, Walmart and startups. The work was production systems and the teams that run them, where a design decision is expensive to undo once it ships.

Today he leads agentic-AI work: multi-agent orchestration, RAG, evals and LLM-Ops. He is also building an agentic-AI product of his own. The judgment he teaches here is current practice rather than a memory of one.

“Anyone can show you the agent pattern. I can show you the three times it was the wrong pattern in production, and how I knew before it cost us.”
— Sunil Mathew

That is what the 30 live hours are for. You bring a design decision from your own work. He shows you how he would examine it, and what evidence would settle it.

10 What you leave with

What physically leaves the room.

Not just capabilities. Artifacts you keep, and your team inherits.

// your system

Your own architecture, pressure-tested

Reviewed with the cohort, with a concrete path to fix what we find.

// toolkit

A reusable toolkit you keep

An evaluation rubric, a threat-model checklist, a reliability/SLO template, and a review-and-governance framework your team can adopt on Monday.

// the room

A small circle of senior peers

And lifetime access to the cohort room, long after the six weeks are done.

11 Live experience

I've spent a career growing people, not just systems.

The craft only counts if it transfers, and it already has. This is the track record I bring into the room.

// mentored

100+ senior engineers, architects & EMs

Mentored and grown across global organisations. These are the people who now run the systems.

// coached

100+ engineers coached

One-on-one and in groups, from strong individual contributors toward real technical leadership.

// taught

~100 senior leaders & directors trained

Ran classes on agentic systems and agentic architectures for senior engineering leaders and directors driving the shift to AI.

// built

Hired & grew engineering teams

Built and scaled high-performing teams across the US, UK, China, and India.

// shipping now

Building an agentic-AI system, as a startup

He is also building an agentic-AI product of his own, so what you learn is current practice rather than a memory of one.

12 The commitment

Make room for the work

The programme includes 30 live hours with Sunil, plus independent work. Confirm the session schedule, independent-work commitment, attendance support and access arrangements before joining.

The cohort is capped at 8 seats. Rolling until all 8 seats are filled. Applying starts a discussion; admission, payment and attendance confirmation are separate steps.

Live hours30, with Sunil
Independent workAdditional; confirmed before joining
Cohort size8 seats, capped
Length6 weeks, live
StartsOctober 2026
AdmissionBy application; every application read personally
CertificateNone
Payment due one week before the cohort starts
Founding seats
Shared on application
The fee is set per region and is shared on application. The first cohort enrols at a founding rate, which rises for the cohorts that follow. Payment is due one week before the cohort starts. Applying commits you to nothing.
  • Starts October 2026 — 6 weeks live, in a founding cohort of 8
  • Your real architecture, pressure-tested live with the cohort
  • Direct access to Sunil throughout
  • Lifetime access to the cohort room
Apply for a seat 

13 Before you apply

Questions before joining

Do I need a company project?

No. Industry-specific cases are the default. Do not share confidential company materials in an application.

Can my employer fund my place?

Yes. Apply as an individual and ask for the employer-funding summary. An enterprise group programme is a separate purchase and scoping conversation.

Is there a certificate?

No. The focus is the capability you develop through practical work and continuous feedback.

How much independent work is involved?

Independent work is additional to the 30 live hours. Its amount and interval will be confirmed before joining.

What are the dates and fees?

Confirm the schedule, fees, payment, refund and access terms in the final offer before committing.

Can we arrange learning for a team?

An enterprise programme normally brings 8–10 participants together with a schedule agreed for the group. Begin with your learning objective, participant experience and industry context.

Is this worth it over a recorded course?

If you want the patterns, a book covers them — they are common knowledge now. This is for the judgment that sits on top of the patterns: live, on your real systems, from someone who has been accountable for the outcome at scale. You are not buying videos. You are buying direct attention and twenty-six years of hard-won judgment.

What does my team get out of it?

The frameworks come home with you: an evaluation rubric, a threat model, a reliability and review-and-governance template. You do not just leave more capable; your team inherits the discipline.

I’m strong but not "staff level." Should I apply?

Apply. Seniority on paper matters less than whether you have shipped something you then had to live with. The application is how that gets read.

14 Apply

Start with your experience and a question

Apply for the open cohort

Tell us about your system-design experience, your current responsibilities and what you want to develop. Let us know whether you are self-funded or exploring employer support.

We use what you write here to reply to you and to prepare the fit conversation. Nothing else. How your details are handled.

Arrange learning for a team

Begin with your learning objective, participant experience and industry context. An enterprise programme is a separate purchase and scoping conversation.

We use what you write here to reply to you and to prepare the fit conversation. Nothing else. How your details are handled.

Ask about the cohort

Ask about the cohort

Ask anything about fit, the commitment, the schedule or employer funding. This is not an application.

We use what you write here to reply to you and to prepare the fit conversation. Nothing else. How your details are handled.

Programme enquiries and employer-funding support, and enterprise enquiries: apply@thelivingcraft.ai