Why we're building Twing

Something changed in the last year that most engineering orgs haven’t fully reckoned with yet: the unit of work an AI agent produces stopped being a function, or even a file, and became a feature. A full PR, touching a dozen files, opened in the time it used to take to write a single one by hand.

That’s a genuine unlock. It’s also the reason a class of problems started showing up that nobody had really had to solve before, because they never came up at this frequency or this speed:

• Two agents, working on the same codebase, quietly duplicating each other’s work — or worse, making contradictory decisions about the same interface, each unaware the other exists.
• PRs arriving faster than any human reviewer can meaningfully read them line by line, which means review either becomes a rubber stamp or becomes the bottleneck that erases the speed gain in the first place.
• All of the reasoning behind why a system looks the way it does — tradeoffs, false starts, hard-won lessons — generated and then discarded in a chat session that nobody will ever read again.

None of these are hypothetical. We ran into all three ourselves, building software with agents, before we ever wrote a line of Twing.

What we’re building

We think three pieces of infrastructure need to exist to make this way of building software actually safe at scale, and we go into each in more depth on our vision page:

1. A collaboration layer so agents (and the humans supervising them) can see what other agents are doing in real time, and conflicting work gets caught before it ships — not reconstructed afterward from a merge conflict.
2. A new review standard, because line-by-line diff review doesn’t scale to this volume. What should travel with a change instead is a record: what the agent did, what it tested, what it touched or weakened in the existing test suite, and what a human actually verified.
3. Compounding organization context — capturing the discussions and decisions that shape a system as they happen, instead of losing them the moment a session ends, so that history compounds instead of evaporating.

The first of these is live today, running in production against real agent sessions. The other two are what we’re building next.

If you’re running into this problem already — multiple agents, one codebase, and no good way to know what they’re stepping on — we’d like to hear about it. Reach us at [email protected].