Hands-on workshop · 40–60 minutes
Build your first useful AI agent.
VibeConf 2026 · 6 October 2026 · Pathé Tuschinski, Amsterdam
Workshop led by Dmitrii Abramov.
Most people begin with “an AI agent for everything”. This workshop starts smaller: one routine that keeps returning, one useful response, and a second-use test that proves whether the agent is worth keeping.
The goal
Success here is not a polished autonomous system. It is a small, testable loop attached to work you already do, with boundaries you can describe and a human decision before anything is kept or acted on.
The workshop is tool-agnostic. Whatever you build in the room should survive a change of platform, because the useful part is the shape of the loop rather than the product it happens to run on.
The method
How do you turn one recurring piece of work into a small AI agent you will actually use again—without building a platform first?
Notice the recurring moment
Start from work that already returns: the same preparation before the same meeting, the same tidy-up at the end of the week. If it only happens once, there is nothing to improve on the second run.
Define the real input
Write down what actually arrives — messy notes, a thread, a list someone pasted. An agent built for the tidy version of the input fails on the day you need it.
Choose one useful response
One output, in a shape you would accept without rewriting it. A summary, a short agenda, a grouped set of observations. Not a dashboard, not five features.
Decide what may be remembered
Say explicitly what carries forward between runs and what does not. Memory is where a small agent becomes useful — and where it quietly becomes a liability.
Keep the human in control
Someone reads the output and decides. Nothing is sent, filed or acted on because the agent felt confident. Every decision should be explainable to the person it affects.
Define the return test
Before you build, name what would make the second use better than the first. That test, not the interface, tells you whether to keep the thing.
What you leave with
A clearly scoped first-agent blueprint and a practical method for building one useful loop: one input, one response, one remembered state, one approval and one return test.
- One routine, scoped narrowly enough to finish.
- A readable first-agent brief you can carry between tools, because it is written in plain language rather than a vendor’s syntax.
- Explicit decisions on memory, control and boundaries, written down rather than assumed.
- A test for the first use and a test for the second.
- One next refinement — not a roadmap.
Depending on the tool and room setup, you may also create a working prototype. The guaranteed outcome is the blueprint and the method—not production software in under an hour.
Examples
These are examples of the kind of routine that works well, not templates you have to copy. Bring your own if you have one.
Before a 1:1
Turn messy notes into a short agenda, and carry forward whatever was not discussed last time.
Weekly operations
Turn scattered updates into a summary someone reviews, with unresolved items kept visible instead of smoothed away.
After customer conversations
Group observations into themes while keeping the source notes attached, so a claim can be checked.
Sales follow-up
Draft next steps from call notes, and require approval before anything reaches the customer.
Hiring and onboarding
Turn repeated questions into a maintained checklist, without putting confidential material anywhere near it.
The smallest complete loop
- One input.
- One useful response.
- One remembered state.
- One human approval.
- One return test.
Tested twice beats polished once.
If one of those five is missing, the loop is not complete yet. With all five present, you have something small enough to judge honestly and useful enough to run again when the routine returns.
Boundaries
- Work with fictional, public or personal-safe material in the room. No confidential, customer or employee data.
- Nothing sends automatically, and nothing takes a consequential action on someone’s behalf.
- Verify outputs before you rely on them. Plausible is not the same as correct.
- Once a prototype is genuinely relied on, bring in engineering, security and privacy. That is the point at which it stops being a personal tool.
Why Dima
Dima Abramov
Founder, Vibe Coding Amsterdam / Product Lead, Foundever
Founder of Vibe Coding Amsterdam and Product Lead at Foundever. Lovable Ambassador (Founding Cohort). He turns ideas into shipped prototypes and runs the hackathons behind this event.
- Founder, Vibe Coding Amsterdam.
- Product Lead, Foundever.
- Lovable Ambassador (Founding Cohort).
- Organises and co-organises hands-on AI meetups and workshops, plus university and European buildathons.
The community archive is the public record of those rooms and the people who built in them.
Keep reading
- Coding with AI for non-developers →What you can genuinely build without an engineering background, and where the limits sit.
- Building products with AI →Define the problem, build the smallest useful loop, and judge the output honestly.
- Vibe Coding Amsterdam →Meetups, workshops and buildathons — and a record of what people built together.
Join us in Amsterdam
Bring one routine. Leave with a loop.
Workshop · 40–60 minutes at VibeConf 2026, alongside the rest of the programme at Pathé Tuschinski in Amsterdam.
€269+ VAT · Regular — Available until 15 September
Get tickets →