The Agentic Engineering Stack

Every agent in this stack is, at its centre, a loop around a language model. So the stack begins by opening that model up.

The course starts from an empty Python file and ends with a 124-million-parameter GPT-2 that you have pretrained and finetuned yourself. On the way you write byte-pair encoding, build attention three times (simplified, causal and multi-head), assemble the transformer block with layer normalisation and shortcut connections, and write a pretraining loop with temperature scaling and top-k sampling. Then you finetune twice: once into a spam classifier and once into an assistant that follows instructions.

Why would an agent engineer need this? Because much of the strange behaviour you meet later has a plain explanation down here. An instruction that the model ignores when it sits in the middle of a long prompt, an answer that changes when the temperature changes, a tool call that arrives with an argument nobody asked for: each of these makes more sense once you have watched next-token prediction happen in your own code.

You will not put your 124-million-parameter model into production. You will understand the models you do put there.

This was one of the first courses we made at Vizuara, and it still does what a first course should do. It puts the whole territory in front of you quickly.

Across ten sessions it moves from what an agent is, to what a large language model is, and then through the frameworks people actually reach for: LangChain and Langflow, smolagents with Langfuse for tracing, multi-agent systems and browser agents, agentic RAG with LlamaIndex, workflows in LangGraph, CrewAI, Google's ADK and n8n. The tenth session is a guest lecture I gave, and the course closes with end-to-end agent projects.

It is not deep, and it does not try to be. Its value is breadth. Afterwards, a job description that mentions LangGraph or CrewAI reads like a list of things you have tried, and you have a sense of which tool suits which problem. The later pieces of the stack take several of these ideas apart and rebuild them from first principles. This is where you meet them first.

A model can act only on what is in its context window. Context engineering is the work of deciding what goes there, in what form, and what is left out. Dr. Sreedath Panat's course is our most complete treatment of it.

The sessions follow the life of a context: foundations, the instructional layer, writing and selecting information with retrieval, tools and MCP, compressing and isolating context, and memory architectures. Then the course turns into building, with a capstone project led by Sreedath, a large build session with me, and a session with Dr. Rajat Dandekar on OpenClaw, reinforcement learning and voice agents. The notes come as PDF booklets, along with a context engineering handbook written by the team.

If the first two pieces teach you what an agent is, this one teaches you why yours does well on the fortieth step of a task and falls apart on the forty-first. Many failures that look like poor reasoning turn out to be context failures: the fact the model needed was never in the window, or it was buried under a page of tool output.

Claude Code is not a model. It is a harness: the loop, the tools, the context machinery and the recovery systems wrapped around a model. In August we ran a five-morning workshop, from 3 to 7 August 2026, on building that layer from scratch. This is the recorded workshop.

Each morning adds one part. The anatomy of a harness and the agent loop. Tool contracts, permission gates and sandboxing. Context budgets, compaction, memory and the CLAUDE.md pattern. Checkpoints, resumable sessions, sub-agents and human-in-the-loop. Finally, production harnesses, with pi, Hermes from Nous Research and Claude Code read as case studies, and a demo of your own harness.

Much of what decides whether an agent is useful happens outside the model: what a tool is allowed to touch, what happens when it fails halfway through, how a session picks up again the next morning. You finish with a working coding-agent harness that you wrote, which is also a concrete thing to show when someone asks what you have built.

An agent that runs only on your laptop is a demo. This course covers the distance between a demo and a product, at a time when much of the code is written with coding agents.

It begins with coding LLMs and the anatomy of coding agents, then builds real applications with them: product requirement documents that drive the agent, MCP servers for reusing integrations, testing and security practices for code you did not type yourself, and deployment with Docker, Terraform, CI/CD pipelines and AWS. It also includes a complete guide to building Claude skills and preparation material for the Claude Certified Architect exam.

It belongs in the stack because the agentic roles being advertised are engineering roles. They ask whether you can deploy, secure and maintain what you build, and this is the piece that prepares you to answer.

A text agent can take its time. A voice agent has a person waiting on the other end, and every stage of the pipeline has to hurry.

Over eight sessions, Dr. Sreedath Panat builds a real-time voice agent from its parts: speech-to-text with Whisper and faster-whisper, text-to-speech with Piper and Coqui, voice activity detection with Silero VAD, the language model as the reasoning layer, tools and memory, and streaming over WebSockets with buffering and barge-in, so the caller can interrupt. The last session covers latency, speech-to-speech models, LiveKit and VAPI, and the course ends with a capstone project.

The reason to learn this from scratch rather than through a platform is the same as in the first piece. When a call feels slow, or the agent talks over the caller, you will know which stage to open.

Formerly AI Agents for Enterprises, and the one piece of the stack that is running live right now. Dr. Rajat Dandekar teaches it on Saturday and Sunday mornings, 7 to 9 AM IST, from 31 October to 6 December 2026: twelve sessions and 24 live hours.

It runs from the transformer and next-token prediction all the way to a deployed agent. Along the way it covers retrieval, a control graph with a certificate gate, and tools that are idempotent and return read-back receipts, so that an agent can be trusted to change a record. Each participant builds a capstone agent for one business module, such as invoice payables, purchase amendments, leave requests, warranty claims or contract review, with tool folders, an MCP server, a harness, a workflow and evals, and demos it live on 6 December against a published rubric.

Inside a company, an agent that chats well is not enough. It has to change the right record exactly once, leave a trail that someone can audit, and stop when it should. This is where the rest of the stack is held to that standard.

On 1 November 2026, DeepSeek Harness published a design note called Exploring Super-agent Design. It describes agents that do not wait for a prompt: a Router that stays on, watches for events and hands work to Workers, with memory that outlives any single session.

In the days after it came out, I explained the note on a whiteboard from start to finish. That video is the last piece of the stack. It is free for everyone, and there is a written version with figures and real-life examples at superagent.vizuara.ai.

It comes last on purpose. Everything before it is about making one agent reliable. This is a look at what happens when agents run all the time, many at once, for a whole team. If you watch it after the other seven pieces, you will recognise every part of it: the model, the context, the tools, the harness and the memory.