Pinned post
Thank you @Gigagoblin for the beautiful review of my talk from Solana Summit Serbia.
What I most want is for devs to feel encouraged to both speed up and slow down with their agents, where appropriate. And most importantly, to keep their work simple.
Full talk below.
What I most want is for devs to feel encouraged to both speed up and slow down with their agents, where appropriate. And most importantly, to keep their work simple.
Full talk below.
14:59
"How to vibe your way to smart contracts ...and some tricks I learned along the way."
Best talk title of Solana Summit Serbia by @pileks, Founding Engineer at @MetaDAOProject, Smart Contracts hold millions in TVL. And then he spent the talk showing exactly how to do that without blowing up.
Some context on why this isn't a gimmick. MetaDAO is the first implementation of futarchy, governance where markets make the decisions, and it has evolved into a full capital markets platform where founders raise in the open. Jure owns the smart contract stack. When your code IS the product and the product custodies real money, "the AI wrote it" is not an excuse anyone accepts. So his workflow has to survive contact with reality.
The core reframe of the talk: vibe coding regular software and vibe coding smart contracts are different sports. A web app bug is a ticket. A smart contract bug is a headline. The environment is adversarial, deployments are effectively permanent, and the entire country of people trying to drain your protocol works weekends. So the question isn't whether agents can write Solana programs. They can, and they write most of his code now. The question is what discipline you wrap around them.
His tricks, and this is the part worth saving:
Plan before you prompt. The agent gets a spec, not a vibe. Writing the plan is where the actual engineering happens; the agent turns decisions into code, it doesn't make the decisions.
Small, scoped tasks. Big open-ended prompts produce confident nonsense. Narrow tasks with clear boundaries produce reviewable diffs.
Tests are the contract with your agent. Let it write exhaustive tests, then make the tests the ground truth. An agent that has to satisfy a hostile test suite is a much better engineer than one being cheered on.
Feed it the right context. Models lag reality, framework versions especially. Give it current docs and your codebase conventions instead of letting it hallucinate an API from two years ago.
Use the agent against itself. Have it attack its own code, hunt for missing signer checks, account validation gaps, and the classic Solana footguns, before a human review that actually reads every line.
Know where the line is. Scaffolding, tests, refactors, docs: go wild. Novel security-critical invariants: slow down, that's where you earn your title.
The honest conclusion: agents made him dramatically faster without lowering the bar, because the bar was never typing speed. It's judgment. The fundamentals, understanding the runtime, the account model, what can go wrong and who profits when it does, matter more in the agent era, not less, because you're now reviewing at the speed you used to write.
As someone who vibe-coded and open-sourced my own analytics dashboard, this whole talk hit home. AI doesn't replace the builder. It removes every excuse not to build.View quoted post on X
