I do wonder what the moat is around this class of products (call it "coding agents").
My intuition is that it's not deep... the differentiating factor is "regular" (non LLM) code which assembles the LLM context and invokes the LLM in a loop.
Claude/Codex have some advantage, because they can RLHF/finetune better than others. But ultimately this is about context assembly and prompting.
There is no moat. It's all prompts. The only potential moat is building your own specialized models using the code your customers send your way I believe.
I think "prompts" are a much richer kind of intellectual property than they are given credit for. Will put in here a pointer to the Odd Lots recent podcast with Noetica AI- a give to get M&A/complex debt/deal terms benchmarker. Noetica CEO said they now have over 1 billion "deal terms" in their database, which is only half a dozen years old. Growing constantly. Over 1 billion different legal points on which a complex financial contract might be structured. Even more than that, the representation of terms in contracts they see can change pretty dramatically quarter to quarter. The industry learns and changes.
The same thing is going to happen with all of the human language artifacts present in the agentic coding universe. Role definitions, skills, agentic loop prompts....the specific language, choice of words, sequence, etc really matters and will continue to evolve really rapidly, and there will be benchmarkers, I am sure of it, because quite a lot of orgs will consider their prompt artifacts to be IP.
I have personally found that a very high precision prompt will mean a smaller model on personal hardware will outperform a lazy prompt given to a foundation model. These word calculators are very very (very) sensitive. There will be gradations of quality among those who drive them best.
The best law firms are the best because they hire the best with (legal) language and are able to retain the reputation and pricing of the best. That is the moat. Same will be the case here.
But the problem is the tight coupling of prompts to the models. The half-life of prompt value is short because the frequency of new models is high, how do you defend a moat that can half (or worse) any day a new model comes out?
You might get an 80% “good enough” prompt easily but then all the differentiation (moat) is in that 20% but that 20% is tied to the model idiosyncrasies, making the moat fragile and volatile.
I think the issue was they (the parent commenter) didn't properly convey and/or did not realize they were arguing for context. Data that is difficult to come by that can be used in a prompt is valuable. Being able to workaround something with clever wording (i.e. prompt) is not a moat.
My intuition is that it's not deep... the differentiating factor is "regular" (non LLM) code which assembles the LLM context and invokes the LLM in a loop.
Claude/Codex have some advantage, because they can RLHF/finetune better than others. But ultimately this is about context assembly and prompting.