G5 Labs raises $14M to make your spec the source code

When agents write most of your code, the code stops being the thing you own. The spec is. On September 15, MIT CSAIL spinout G5 Labs came out of stealth with $14 million in seed funding on exactly that bet: write down what the software is supposed to do, in structured natural language, and let that artifact — not the repo — be the source of truth. Whether or not you ever buy the product, the argument is one every team shipping agent-written code should sit with.

What actually happened

VentureBeat reports that G5 Labs, founded by MIT computer science professor Tim Kraska, raised $14M co-led by Pillar VC and Battery Ventures, with Omega Venture Partners and Encoded Ventures joining.

The platform compiles business requirements and organizational policies into what the company calls a system ontology — a structured semantic graph. The compilation runs both directions: existing code gets lifted into a semantic representation, and updated specifications regenerate implementation. The claimed payoff is conflict detection at the level of meaning rather than text — two agents writing changes that both merge cleanly and still contradict each other is a problem git was never designed to catch. G5 sits above the coding models rather than replacing them, working over Claude Code and Codex.

Kraska's framing: natural language, with structure on top, becomes the new source code. The context he cites is the scale problem — Anthropic reports AI now authors up to 80% of its production code, and VentureBeat notes one engineer submitting a 300,000-line change in a single week. Around 90% of G5's early customers are in financial services. Customer names are under NDA, and the company's claims about lifting millions of lines of legacy code into ontologies are self-reported, so weigh them accordingly.

Why the spec matters for your business

You are not buying an ontology compiler. You are about to have the problem it exists to solve, at one-thousandth the scale.

Here is how it shows up in a five-person shop. An agent adds a discount rule to the checkout. Three weeks later a different agent, in a different session, adds a loyalty credit. Both pass tests. Both ship. Then a customer stacks them and pays $0.40 for a $90 order — and nowhere in your repo is there a sentence saying these two things may not combine. The rule existed. It lived in someone's head.

The fix is unglamorous and it is free: keep a written, versioned specification of the rules your system enforces, in the repo, in plain language. Pricing rules. Refund limits. Who can see what. What the system must never do. Review it in the pull request alongside the diff. When an agent proposes a change that contradicts it, that's a conflict — and you caught it by reading, not by getting a support ticket.

That file is also the thing that survives your model choices. Models get deprecated, priced up, and swapped out. A specification you own in your own repository ports to whatever writes the code next quarter. The generated implementation is increasingly disposable. The intent behind it is not.

Sources: VentureBeat, Unite.AI.