Seven levels of determinism.
From a sentence in a prompt, through skills, tool permissions, subagents and structured MCP tools, to hooks and CI. A one-page decision card tells you where any agent rule belongs.
Hands-on workshop 100% free
Guardrails, not guidelines.
Turn your requests into guarantees.
Most AI coding agents grow one rule at a time, until they run on a long instruction file and still push to main on a bad day. Join Kuba Wysocki to learn the determinism ladder: where each rule belongs so it holds when nobody is watching.
No fee, no ticket needed. Register with your e-mail and we will send you the link to join. With Kuba Wysocki · software engineer at Conscensia, author of ctxRunner and BitByteBench. Tool-agnostic: Claude Code, GitHub Copilot, Cursor or similar.
RULES THAT HOLD
Kuba WysockiSoftware engineer at Conscensia · ctxRunner, BitByteBench
From chaos to control
A rule you shout.
A gate that stops.
Rules pile up around the agent, and it still reaches for "Push to main". Move the rule into a gate, and the same action simply does not happen. Then the agent takes the allowed path, runs the tests and gets through.
PromptPermissionsHooksCI
What we'll use
Determinism ladderHooksMCPSubagentsCI
Registration Free
Save your seat.
It's free.
About 70 minutes, hands-on, with Kuba Wysocki: the determinism ladder for AI coding agents, a live hook stopping an agent mid-action, real measurements, and your own refactor of a deliberately bloated agent. Online, no fee, no conference ticket needed.
Add to calendar
Free registration
Enter your e-mail below. That's all.
- You get the link to join the live session by e-mail.
- We remind you shortly before we start.
- No payment, no card, no ticket.
Loading the registration form…
Guardrails that hold.
Tool allow-lists, deterministic context through MCP, hooks, subagents with contracts and CI as the backstop. Plus what should stay as plain prose, what to delete, and an honest list of what these mechanisms do not fix.
Your own 20-rule agent, fixed.
Refactor a deliberately bloated agent yourself, watch a hook stop an agent mid-action live, and see real measurements of a prose-heavy agent against a focused setup on the same task.
What it's about
Writing the rule louder doesn't fix it. Moving it does.
Most AI coding agents grow the same way. Something goes wrong, someone adds a rule. A year later the agent runs on one long instruction file: hundreds of rules, half of them in capital letters, re-sent with every request. And on a bad day the agent still pushes to main, skips the tests or opens the secrets file, then reports that it followed every rule.
This workshop introduces the determinism ladder: seven levels from a sentence in a prompt, through skills, tool permissions, subagents and structured MCP tools, to hooks and CI. You'll learn how to decide where each rule belongs, then refactor a deliberately bloated 20-rule agent yourself. You'll see a live demo of a hook stopping an agent mid-action, and real measurements from a demo repository comparing a single prose-heavy agent with a focused setup on the same task.
We'll also look at what this means for teams in larger organizations: compliance you can prove, guardrails that hold when nobody is watching, and a way out of slow AI governance that reviews documents instead of mechanisms.
You'll leave with
- A one-page decision card for where any agent rule belongs
- Concrete patterns: tool allow-lists, deterministic context through MCP, hooks, subagents with contracts, CI as the backstop
- A sense of what should stay as plain prose, and what to delete
- An honest list of what these mechanisms don't fix
- The open-source demo repository from the talk: both agent setups side by side, with hooks, tests and CI, running on macOS, Linux and Windows
Who it's for
Developers, tech leads and platform or DevEx engineers who use or roll out AI coding agents (Claude Code, GitHub Copilot, Cursor or similar). The concepts are tool-agnostic. Some hands-on experience with an AI coding agent helps.
What to bring
Nothing to install. Keep the chat open, and have a pen or a notes app ready for the exercise. If you have an agent instruction file of your own, bring it: the last task is to move one of its rules. To explore the demo repository afterwards you'll need Node and the .NET SDK; Claude Code is optional.
About 70 minutes. Hands-on.
From a rule you shout
to a gate that holds.
We start from the instruction file every agent accumulates, climb the determinism ladder, watch a hook do its job, and finish by moving a rule of your own.
Guardrails, Not Guidelines- 01
The instruction file problem
Hundreds of rules, re-sent with every request, and an agent that still pushes to main and reports that it followed them all.
- 02
The determinism ladder
Seven levels, from a sentence in a prompt through skills, permissions, subagents and MCP tools to hooks and CI. How to decide where a rule belongs.
- 03
Live: a hook stops the agent
An agent is caught mid-action by a hook. The difference between a request and a working guardrail, in one minute.
- 04
Measurements, not opinions
A prose-heavy agent and a focused setup on the same task in the demo repository, with the model held fixed.
- 05
Hands-on: refactor the 20-rule agent
You sort the rules onto the ladder, decide what stays as prose and what gets deleted, then move one rule from your own instruction file.
- 06
Teams, governance and Q&A
Compliance you can prove, guardrails that hold when nobody is watching, what these mechanisms do not fix, and your questions.
Meet your host
Kuba Wysocki
Software engineer at Conscensia · author of ctxRunner and BitByteBench
Kuba Wysocki is a software engineer at Conscensia, where he builds .NET backend services for a large multi-tenant SaaS platform and helps roll out an AI-first way of working: agent conventions, guardrails and the tooling that turns AI-assisted development from an experiment into something a team can rely on.
Learning C# is when he fell for programming. AI-assisted development turned out to be the second wind he didn't know he was waiting for, and AI itself is the Alfred Pennyworth he always deserved: always on call, never tired, occasionally a little too eager to help. Even Alfred works better with house rules. And no, "You are an expert", "Be concise" and "Don't make mistakes" are not house rules. Those are magical incantations we whisper, hoping the day we have to read the actual documentation never comes.
He took the scenic route into software: a degree in emergency medical services, a stint in anti-money-laundering at a bank, then years of QA and test automation before moving to C# full time. That probably explains his weakness for checklists and guardrails. These days he builds ctxRunner, a tool that gives coding agents deterministic, reviewable context, and BitByteBench, a benchmark that compares agentic setups on the same task with the model held fixed.
Good to know
Before you join.
Is the workshop free?
Yes, completely. There is no fee and you do not need a conference ticket. Just register with your e-mail.
What will we do?
Learn the determinism ladder for AI coding agents, watch a hook stop an agent mid-action, look at real measurements, and refactor a deliberately bloated 20-rule agent yourself. You leave with a one-page decision card and the open-source demo repository.
Do I need to install anything?
No. Keep the chat open and have a pen or a notes app ready. If you have an agent instruction file of your own, bring it. To explore the demo repository afterwards you will need Node and the .NET SDK; Claude Code is optional.
Which tools does it apply to?
The concepts are tool-agnostic: Claude Code, GitHub Copilot, Cursor or similar. The live demo and the repository use one setup, the patterns transfer to the others.
Who is it for?
Developers, tech leads and platform or DevEx engineers who use or roll out AI coding agents. Some hands-on experience with an agent helps.
When does it take place?
The workshop takes place on Wednesday, 4 November 2026 at 17:00 CET (Europe/Warsaw). The live session lasts about 70 minutes and is held in English.
How do I join?
Register with your e-mail in the form above. You will receive the link to the live session by e-mail, plus a reminder before we start.
Ready? Free
Save your free seat for Wednesday, 4 November.
Takes 10 seconds. You only need an e-mail address.
Save my free seat