A live, hands-on implementation masterclass for AI engineers, developers, solution architects, and GenAI practitioners — 120 minutes, start to finish. We build an agent harness from scratch in plain Python, deliberately break it the five ways real harnesses degrade, and fix each failure in front of you.
Book a Seat
Tell us a bit about what you're building and we'll confirm your seat within 24 hours. Register now and we'll email you the moment the next date is confirmed.
Free · 120 minutes · next date announced soon
By registering you agree to receive the live session details. We'll never sell your data.
The quickstart question: “How do I get an agent calling tools?”
The production question: “Why is it behaving differently than it did last week?”
Getting a loop to call a tool takes about twenty lines. What takes weeks is everything the harness decides on your behalf: what gets dropped when the context fills, what happens to a tool call that fails, how many times it retries before giving up, which model is actually running, and what that tool layer can genuinely reach.
This masterclass is about the layer you own.
Built For You If
You're an AI engineer, software developer, solution architect, technical lead, or GenAI practitioner
You've got an agent running — framework or homegrown — and you're not confident it survives a month of real use
You're putting an autonomous loop somewhere it can touch real files, real credentials, or real customers
You can read and write Python, and you'd rather see a system built than watch slides about one
You're the person who gets asked why the agent 'got worse', and needs a better answer than a shrug
Not For You If
We don't stop at the quickstart. We start where it ends.
Why this is different
Almost everything published about building agents stops at Tier 1 or Tier 2. We start at Tier 3.
Twenty lines, a while-loop, a tool call, and something that answers. Genuinely useful for seeing the shape of the thing — and where almost all agent content stops.
A walkthrough of one library's abstractions exactly as advertised. Fine until the loop behaves differently in week three and the tour never covered what the abstraction was deciding for you.
What you need after the loop works: a context budget you set, tool errors the loop cannot swallow, a retry ceiling, a model you can prove, and a blast radius you chose. For engineers running agents against something that matters.
Outcomes
Five concrete things you walk away with — designed to be applied to your own AI application on Monday morning.
A working agent loop in plain Python — context assembly, tool dispatch, result handling, and a stop condition — built in front of you rather than imported. No framework, so there is nothing in it you cannot explain or change.
Context rot, tool calls that fail silently, a retry loop that never converges, a model that changed underneath you, and permissions wider than you assumed. You'll recognise each from the symptom instead of guessing for a week.
Context management, tool mediation, failure handling, and termination are four separate jobs that fail differently. Which one to inspect first, what each log line really tells you, and why fixing the wrong layer makes the problem harder to find.
What your tool layer genuinely exposes, why an allow-list is not automatically a boundary, and how to scope credentials, filesystem and network access before the agent has a bad day rather than after.
A practical pre-launch checklist covering context budgets, tool error contracts, retry limits, model pinning, instrumentation, and recovery. Written to be applied to your own harness on Monday morning.
Live Implementation · Not Slides
Once the harness works, we break it — the same five ways real agent loops degrade in production. Each one gets reproduced, diagnosed, and fixed in front of you, with the reasoning spelled out.
Failure 1
The window fills, something has to go, and the harness decides what. We watch an agent lose the one constraint that mattered, look at what truncation and summarisation each actually cost, and tell compaction loss apart from a model that simply got the answer wrong.
Failure 2
A malformed argument, a stack trace handed back as if it were a result, and an agent that reasons confidently on top of a failure it never noticed. We reproduce it, then give the tool layer an error contract the loop cannot ignore.
Failure 3
The same failing call, eleven times, at full price. We watch it happen, then add a real budget — attempt limits, backoff, and a give-up path that escalates instead of spinning, because 'stop and ask' has to be a first-class outcome.
Failure 4
A fallback fired, a version moved, a default changed — and the loop is running something weaker than you configured. It looks exactly like the harness getting worse. We reproduce it, then make the running model impossible to be wrong about.
Failure 5
The failure underneath the other four. What the tool layer can genuinely reach, why the allow-list you wrote is not the boundary you pictured, and how to make the blast radius a decision rather than an accident.
Nothing to install beforehand — we build everything live · What you'll see in the stack
Why Learn From Me
AI Consultant & Architect, IdeaToMVP · 12+ years in AI & enterprise systems
linkedin.com/in/surya-pratap-singh095I've trained engineering teams at TCS, IBM, and other enterprises on production AI — agents, RAG, and MCP — and run industry workshops for teams shipping AI into real products. As a consultant, I've helped companies rearchitect how they deploy AI, shipped 15+ systems to production, and trained 300+ professionals.
Every one of the five failures in this session is one I've debugged in a live system — usually at the point where the demo had already convinced everyone it worked, and usually in the harness rather than the model. This masterclass is the session I wish existed before those debugging weeks.
Agenda
Tight, rehearsed, and ending exactly when promised. Questions are welcome throughout, not just at the end.
The gap between an agent that answers once and one you'd leave running. The four jobs a harness is really doing, and the five failures we're going to cause on purpose.
From an empty file: the loop, context assembly, tool schemas and dispatch, result handling, and a stop condition you chose — in plain Python, no framework, every decision visible.
Context rot, a silently failing tool call, a retry loop that never converges, a silent model change, and permissions wider than you assumed. Reproduced, diagnosed, and fixed in front of you.
What to log so 'why did it do that' has an answer, which signal tells you which layer broke, how to bring a degraded loop back — and the production checklist you take with you.
Bring the harness problem you're actually stuck on. Questions are welcome throughout, and we close on the hardest ones.
FAQ
Yes — this is an implementation session. You should be comfortable reading and writing Python and following along in a terminal. You don't need to have built a harness before; we write it from an empty file.
No. We write everything live, so you can watch or follow along, whichever you prefer. If you want to build along, have an editor open, a terminal ready, and access to a model — local or an API key, either works.
None, deliberately. We write the harness in plain Python against a model API so every decision is visible rather than made for you by an abstraction. The failure modes are architectural, so everything we cover maps onto LangChain, LangGraph, the Agents SDK, a coding agent's harness, or whatever you've grown in-house.
That's the ideal case. The build is the warm-up: its point is to make visible what your framework is already deciding for you. The value is the five failure modes and the diagnostic order of operations. Bring the behaviour you're seeing and we'll place it in the Q&A.
Yes — recording is sent to all confirmed attendees. The live Q&A is reserved for those who attend live.
None, and no. The session is free. At the end I'll mention the work we do at IdeaToMVP for teams who want help shipping this, but you're under zero obligation. The 90 minutes stand alone.
Registration takes under a minute. We confirm seats within 24 hours.
Book a Seat →