Research
Current AI agents have enabled a latent skill that was perhaps dormant in many a curios people, but might've awakened when computational power became leveragable via natural language (a.k.a ambiguous thoughts). This prototyping and quick POCs, have given rise to a lot of fun and truly creative ideations, which previously were too far away from a generalist.
NOTE : One simple lesson I learned that even though I can explore various avenues, without at least some domain knowledge, I get pretty lost in these POCs/Prototypes/research ideas soon after entering them. Jargon words are thrown by LLMs, deciphering which takes more tokens than nought. Although the model can suggest and test ideas on it's own, I cannot back them and waste precious tokens on ideas I am mostly unfamiliar with. Given this, below I'll both list the ideas I thought of exploring, my inspirations and the main bottleneck (learning more about the domain) that made me pause, and deliberately choose my next step forward.
There's a central skill here that I'll develop nonetheless, and would like to apply it in various places (can the LLMs do it better then me ? I think they already are. But it still pays to be a domain-whisperer, in a world where LLMs already are domain-wizards). That central skill is proof engineering. This is very close to what ProofCraft offers, but at a very lower level, as I am but a novice. So with that out of the way:
Provide formal methods and verification to business workflows. I have been toying with this idea ever since I learnt of Dependant types, where we perform very strict type checking, to the point of mathematical guarantees on certain business workflows. This may have some practical implications, as certain business (especially the ones dealing with critical systems) can really benefit from faultless systems and workflows. This seems like the closest one for me to implement at least a prototype for (although still a non-trivial task) (This is already being done, in cases like the various cases of TLA+ usage, MicroSoft's Project Everest or other Static Analyzers for aviation). This process has traditionally been very expensive and tedious, which is the problem being solved by the agents. Although deciding and constructing a proof might still require some knowledge, which I believe people need not spend decades on learning, the rest can be handled by the agents.
Mostly realized by AlphAxiv's Autoresearch, but performing small research has never been easier and more accessible. Though once again domain knowledge is required, a person with enough curiosity and an agent can implement and explore various "paths", alongside implementing and graphing them. This has become a very interesting way to quickly mix and match certain ideas, try methodologies on certain concepts vs others etc ... Various ideas may pop-up in our heads during reading about certain research, and we no longer have to wait to explore most of our ideas (which can sooner get us to the realization that good ideas are extremely rare :)). Some examples that I love are: this author who explored how to improve models without training them, this person who realized programs can be SQLite databases and the high profile undertakings of bun re-write in Rust. Larger re-writes and faster prototypes are now possible and may lead to great software (given the agents are actually paired with curious, clear headed and driven individuals ;)
Cross domain offerings. This, for me seems to be the most interesting application, as it involves both the programmers and the domain-experts to learn and grow using the new tool in town. A world where simply writing code is not valuable anymore, we must alternatives. Thankfully, since SWE jobs (except for very entry-level jobs) was never about just typing code, we may survive for a bit more, the beginning of a "computational generalist" is here. The idea of computation can be applied to various other domains, which I have either tried to explore or stumbled upon. Here is where I felt that simply knowing the computational side of it all wasn't enough, as LLMs, in all their convincing arguments, could not stop me from feeling uneasy stepping into unknown territory, we must learn and adapt. Syntax never mattered, what mattered was understanding the idea of computation and in modern times, learning how to leverage it. One example is Proposogrphy where we leverage graph theory and probably graph DBs to History or this example of how any data can be peeked into by outsiders (non-domain experts), and taken advantage of.
Curiosity is innately human. The need to learn and explore is what makes us feel human, be it in the form of art, writing or research. Agents as we know them are a giant step towards innovative research, and it'll be a waste if we do not see it as such. Stay curios, and leverage !