The mission is straightforward: put our best in class coding agent behind a task orchestration agent we call cto.new, make it free and empower a generation of software engineers to break through the clouds and soar into wild new levels of productivity.

Sorry, what?

Feels like we’re already somewhat carried away here, so let’s back up a little to see how such a big statement is actually predicated on a couple of very clear learnings and real ambitions. 

Why an orchestration agent?

Relatively speaking, the coding agent is feature-complete. Let me qualify that: we’re seeing our users actually merge over 60% of all pull requests created by our agent, and if you take a full helicopter view of our entire platform usage, the cost for these is a couple of dollars each. Pretty strong signal.

We see, however, the coding agent alone as a much smaller fraction of the bigger picture.

From assembly to scripting languages, productivity has always risen by offloading work from humans to machines - LLMs now take that much further, letting developers operate at an unprecedented scale.

There’s a huge opportunity space in which to draw the rest of the owl here and develop a wholly new software engineering experience with people in the driving seat. The current cto.new is the very first, nascent step in what we envisage will be a big journey to solve for the core problems and challenges of empowering people in this new space.

Why free?

Naturally we’re fundamentally, passionately bullish on the next layer of abstraction for software development, and this belief fuels our genuine desire for empowering access for all.

The extreme rate of change in this space makes time by far the most vital asset along this journey. Making a free product is simply the fastest way to get there, developing in partnership with real teams, codebases and their traces.

Those things out of the way, let's get back to the knife juggling part.

Knife juggling

It’s no surprise that putting a free product out in this space comes with a hefty side order of distracting challenges that come under the umbrella of “fair usage.” 

We have the expected abusive use, which we’re not covering here, and some challenges with pure slop users where the ROI on high token usage is clearly never going to work.

Those aside, more interesting is our philosophy for legitimate users, which comes down to a subtle re-framing of the classic principle of least access to a dynamically positive assumption that users will interact well. AKA, trust but verify. Our intent is for legitimate users, engaging with the app as intended, not to experience rate limiting. 

This manifests as a growing number of heuristics, continuously internally evaluated, that allow us to silently reward the best behaviours while imposing some guidance and, ultimately, sanctions where needed.

We’ve given ourselves three simple levers:

1. Dollar cost - we do the auditing work to handle multiple token counting systems to ensure we always know the current rate of spend per user.

2. Rate limit ceiling - enables correlation of high quality interactions with unencumbered access

3. Rate limit time window - enables some control of severity and hence guidance, from sin bin timeouts to more aggressive blocking

Rules themselves vary from simple user interaction counts that put a measure on engagement to more sophisticated agent classifiers that respond async to traces, and keeping this logic entirely server-side helps us prevent subversive gaming of these rules.

There’s a long way ahead, and a great deal of fine tuning to handle as we grow, but we’re hopeful this reward based system will help us keep the lights on whilst we tackle the big product challenge.

We’d love you to try cto.new for yourself and let us know what you think in our Discord community. Thanks for reading!