vdaubry

Engineer enjoying AI psychosis
🎖️ CTO & co-founder at Brava
⌛ Prev. founder @nextoryfr

[vdaubry @vdaubry]

We shipped our 1000th user story last week with our internal agent orchestration tool. To celebrate, I'm open-sourcing it 🎉

For the past 4 months, 100% of our production code has been written with this tool

Remote first, Multi-user, Multi-Harness (Claude, Codex, OpenCode, etc)

[vdaubry @vdaubry]

@dadadaistt Yes that's our biggest lesson :)

Where should we invest our time now that code is produced in the background ?
Reviewing PR was the bottleneck.

We shifted our focus on the plan, and designed a process that minimizes human intervention between the plan and production.

[vdaubry @vdaubry]

We shipped our 1000th user story last week with our internal agent orchestration tool. To celebrate, I'm open-sourcing it 🎉

For the past 4 months, 100% of our production code has been written with this tool

Remote first, Multi-user, Multi-Harness (Claude, Codex, OpenCode, etc)

[vdaubry @vdaubry]

@maxfaber_Om 2) Models are still sycophantic :)
I don't mind, the main issue I wanted to solve is what comes after the plan : merging PRs fast (PR matches the plan), and reduce bugs in production (agent must perform manual QA on the PR)

[vdaubry @vdaubry]

@maxfaber_Om 1) yes. The inital plan is rarely perfect.
Planning phase is super interactive, I treat the agent as a co-worker, we discuss the plan until we come to an agreement.
Once the plan is solid, and I click "start", changes are deployed to production with minimal human intervention.

[vdaubry @vdaubry]

@maxfaber_Om Haha, that was a fun experiment !
August 2025 was around the time we switched to 100% of the code written by agent.
At the time it was the lot of pain... Sounds like ancient history.

This lead us to build vdaubry/bottega ;)

[vdaubry @vdaubry]

A fun experiment we conducted to improve an agent’s ability to self-evaluate its front-end work

https://t.co/svpBXcdoTq

[vdaubry @vdaubry]

@maxfaber_Om I checked MaksimZinovev/docfence, this is definitely something that could help with the planning phase !

I spend 50%+ of my time on getting the plan right, this is where most of gains lies.

The PR quality issue is more or less solved for us now (>80% approval after review)

[vdaubry @vdaubry]

Fork it, break it, build your own version.

If you're building something similar, or you disagree with any of this, I'd love to hear from you.

Repo: https://t.co/cxeVfHq2hf
Writeup: https://t.co/cvlxmvJ0S3

[vdaubry @vdaubry]

The loop just makes the agent behave like a human dev: plan → implement + tests → an adversarial review agent that re-runs the test scenarios and checks no checkbox was silently skipped → runs QA on the final task via MCP -> open the PR, keep CI green.

The human owns the plan and the review.

[vdaubry @vdaubry]

Karpathy's autoresearch: give an AI agent a codebase, let it run experiments overnight, wake up to results.

Generalizes to anything — load testing, landing page A/B tests, infra tuning...

"Program the program" is the real insight, I love it :D

https://t.co/aSL8tivHnj

[vdaubry @vdaubry]

@nlw @systemofadown @Korn 🤘

[vdaubry @vdaubry]

Dive into the era where language is just a tool! Discover why with AI, even unsexy languages empower solo experts to build big. Curious how? Find out more!

https://t.co/1R39i1SgFH

[vdaubry @vdaubry]

Unlock AI's true potential by breaking big tasks into bite-sized prompts! Discover why detailed instructions lead to better automation results. Dive in and elevate your AI game!

https://t.co/NY2KAwkiMv

[vdaubry @vdaubry]

Why avoid writing small bits of code when they can save you hours? Discover how the cost of coding is dropping and boost your productivity!

https://t.co/XjgMTeTOiJ

[vdaubry @vdaubry]

@po_petit This place is amazing! a beautiful walk indeed :)

[vdaubry @vdaubry]

@po_petit @FonPernodRicard c'était puissant ?