A rise of productivity… but only consistent after a while.

AI agents (for coding or even for more complex tasks) are now part of the routine of most – if not all – software developers in the world. This is true for me too, and it significantly speeds up Arcsecond’s roadmap. 

For a while, the productivity increase in coding with AI was overwhelming and even disturbing, because it was opening too many possibilities at the same time. After a few weeks, a few key elements became clear, however.

First, the feeling of acceleration was particularly true because Arcsecond codebases were rigorously and extremely well structured before the introduction of AI. Such code organisation and consistency made very easy for AI agents to follow existing patterns, and further build on existing decisions.

Second, the fraction that remains purely-human in this new work mode becomes even more important. And it is the product vision. Since the cost of implementation of new features is now close to zero, what remains is the consistency of the vision of the software over time.

This vision obviously benefits from zero-cost exploratory developments. But, in our opinion, it remains critical for developing a product that is designed for the coming decades in mind, supporting real-life operations. 

Is Anthropic’s “Fable 5” model capable of cloning Arcsecond in a few days / hours?

Since a few weeks, the acceleration of AI capabilities is even stronger, and there is no reason to think it will slow down anytime soon (as of today, “Fable 5” of Anthropic has just been released). The questions we are facing now are even bigger. AI coding agents may be so powerful that it is tempting to think that even complex software like Arcsecond (whose codebase started over 10 years ago) could in principle be “easily” reproduced with well-configured agents, a good harness, enough time and tokens. 

Indeed, that’s a possibility.

But again, a similar pattern emerges. What is left to humans is the thinking they put inside “irreducibly complex systems” (quote from Helen Edwards – The Artificiality Institute).

Safely running an entire observatory involving multiple people, permissions and hardware is not something that is easy to describe. And a responsibility we do take extremely seriously. And even if you have the experience to describe it “all”, real-life operations are still the best and only true validation, in order to slowly incorporate heuristics into the software system. Not even mentionning the purely-technological issues that still require a deep knowledge and experience of an operating software in specific production environments. 

Conclusion

All in all, I adopt a reasonably enthusiastic approach. Trying to get the best of it, yet keeping a truly critical eye on the result. For our roadmap, it means that our Night Scheduler, along with the new Safety Procedure Manager tool, will be made available and tested in real conditions later in August this year, with a mission to Tasmania (opening the development window of the Queue Mode for the end of the year). 

We expect that by September, at least two observatories are running with Arcsecond for their operations.

P.S. This post has been generated only with the usual sweat and tears of a human…