Google × Kaggle · 2026
Winner of the Agents for Business track
Five days of the Google × Kaggle AI Agents Intensive, one capstone project, and four lessons I now apply in every client engagement. With open code, a public video and a live demo.
The capstone: an estimating agent for an electrical contractor
I did not take a textbook problem. A contractor receives a sketch of a site — often a photo of a hand-drawn plan — and spends two to three days turning it into an estimate, a price offer, a contract and a shop order. The capstone does that path in about ten minutes: a sketch, a PDF, a voice note or text goes in; a document package and a take-it-or-leave-it margin verdict come out.
Inside are three agents talking over the open A2A (Agent2Agent) protocol, taking prices from a shared MCP server on top of a real price list. Every model call is wrapped in guardrails: a prompt-injection filter in two languages, a budget ceiling, and output validation. The version published on GitHub is fully anonymised — no client data in it.
Project analyst agent
the only component that calls a model
Reads a photo of a sketch, a PDF, a voice note or plain text through computer vision and turns it into a structured project. When the input is thin it asks the human a clarifying question instead of inventing scope.
Document factory
deterministic code
Assembles the price offer, the contract, the invoice and the shop order. No model involved: the same input produces the same output, every time.
Margin advisor
deterministic code
Computes the margin, works backwards to «how many crew-days fit inside the target margin», and returns a verdict: take the job or walk away.
Four lessons I took into client work
An agent where judgement is needed. A tool where it is not.
The model sits at exactly one point in this system: reading unstructured input. Everything that touches money is computed by deterministic code. In a task that must be right to the last unit, an LLM doing the arithmetic would be both more expensive and worse.
Computer vision is non-deterministic. Process fixes that, not prompts.
The same photo of a sketch produced a different scope of work from run to run. The fix was structural: a strict extraction schema, clarifying questions instead of guesses, and a human who confirms the estimate. The AI drafts; the person decides.
Test the agent on real input, not on synthetic data.
The price matcher worked perfectly on test fixtures and then silently returned zero prices on real computer-vision output. An agent that reads images has to be evaluated on images of realistic quality — anything else measures the wrong thing.
A specification written before the code pays for itself.
The behavioural spec was written before the first line of code and still describes the system: «the agent must ask, not invent», «an unknown line item is flagged, not estimated». It survived every refactor because it described behaviour, not implementation.
Honestly, the most useful part of the course was not the technology but the framing. Before it, my question was «what can an agent do?». After it: «where exactly in this process is judgement needed, and where is ordinary code enough?». Answering the second question saves the client money twice — on tokens and on mistakes.
What the course covered
The «5-Day AI Agents Intensive» is a free course from Google and Kaggle: a whitepaper, a podcast, a live stream and hands-on code for each day. It ends with a capstone project that goes through review — the certificate is awarded for the project, not for watching the lectures.
Agentic development
An agent is a model plus its scaffolding. Context engineering as the core skill. How evaluating an agent's trajectory differs from testing code.
Tools and protocols
The open standards agent systems are assembled from: MCP for connecting models to tools and data, A2A for agents talking to each other.
Agent skills
A skill as an agent's procedural memory: a folder with instructions, three levels of disclosure, composition, and the Read / Draft / Act ladder of permission.
Security and evaluation
Context as the security perimeter, sandboxes, prompt-injection defence, observability, and evaluating an agent across seven dimensions.
Specifications and AgentOps
The spec as the source of truth (BDD/Gherkin), zero-trust, shipping safely: roughly 80% of the work on an agent is the scaffolding, not the model.
See it for yourself
The demo runs within a daily request limit — if it is rate-limited when you arrive, the four-minute video shows the same path end to end.
Let’s talkabout your bottleneck
Hiring, a project, or a second opinion on an agent design — Telegram is the fastest way to reach me. I read everything myself.
mail [at] artemborisov.ru
Based in Russia (UTC+12) · open to relocation