Inspiration It started with a smoke test. I was working on a project and asked a brand-new, top-of-the-line model to do something that should have been trivial — get a basic smoke test to pass — and it simply couldn't. Not "did it imperfectly." Couldn't. So I tried others. Same story across the board: even the most recent, most capable models were bad at smoke tests, almost universally.
That stuck with me. If the smartest models we have can't reliably clear the lowest bar of "does this actually run and pass its own checks," the problem isn't the model's intelligence — it's how we're asking it to work. My whole goal became narrow and stubborn: make it so an AI can complete a smoke test as reliably as a human can. Everything else grew out of chasing that one thing.
What it does Chasing reliable smoke tests is what led me to intent mapping — the realization that the failures weren't coming from bad typing, they were coming from unstated intent. The model was deciding and coding at the same time, so nobody — not even the model — knew what it had actually committed to.
So Code 2.0 separates the two. Your intent is captured as an explicit, machine-checkable blueprint before any code is written. A cheap model transcribes that blueprint into working code. Then a deterministic harness — not an AI — runs every acceptance and end-to-end check itself and issues a cryptographic receipt proving the build passed. The coder never grades its own work. Green, or it doesn't ship.
How we built it The engine is JSON schema at its core — intent gets mapped into a strict, validated blueprint — refined over several rounds of deep research until the system clicked into place.
The logical progression that got me here:
Blind coding → capture the intent → mapped intent becomes a task DAG (Directed Acyclic Graph) → a DAG can be verified, pre-checked, and split → splitting invites many agents → many agents must stay in accord with the plan → Concordance → strip the parallelism (couldn't get it working yet — and it was blocking me from shipping) → Code 2.0.
Code 2.0 is the point on that path where the product was already complete — mapped intent, verified code, a receipt — before the DAG ever needed to fan out to a swarm.
Challenges we ran into I got distracted by agentic parallelization. The DAG begs to be run in parallel, and I chased that — building toward a governed multi-agent system where every agent stays in concord with one plan. It was the most ambitious version of the idea. It was also the thing standing between me and a shippable product. Recognizing that, and cutting it, was the hardest call I made.
Accomplishments that I'm proud of The blueprinting system — turning fuzzy intent into a formal, checkable contract.
What's next for Code 2.0? Agent-parallel coding. The parallelization I stripped out to ship isn't dead — it's the roadmap. Code 2.0 proved the single-agent product works and pays. Concordance is what it becomes when the plan governs not one agent, but a whole swarm acting in concord.
Log in or sign up for Devpost to join the conversation.