Your third prediction about the domain knowledge problem is the one I keep coming back to. You're right that the models are smart enough but lack the context.
I reckon the answer isn't a single new technique but more of a pattern. You feed the agent your team's conventions, architectural decisions, and coding standards through structured instruction files. Not just longer context windows but curated context that tells the agent how your codebase works. I wrote about this with OpenCode agents specifically https://blog.devgenius.io/your-senior-devs-dont-scale-your-opencode-agents-can-e2ecf2d04548 and found it eliminates most of the "technically correct but doesn't fit" output that makes people distrust agent-written code.
Basically your prediction might already have a partial answer in the form of agents-as-code, where the domain knowledge lives in config files that travel with the repo. Did you end up seeing anything along those lines since?
We've tried variants of that (for Claude Code) and it feels hard to get right. Right, meaning, CC will reliably follow this advice and not just ignore it half the time. I think maybe there's a skill issue in how we are writing the instructions.
But I'm also curious how this works outside of coding, in fields that are less defined by plain text files.
Your third prediction about the domain knowledge problem is the one I keep coming back to. You're right that the models are smart enough but lack the context.
I reckon the answer isn't a single new technique but more of a pattern. You feed the agent your team's conventions, architectural decisions, and coding standards through structured instruction files. Not just longer context windows but curated context that tells the agent how your codebase works. I wrote about this with OpenCode agents specifically https://blog.devgenius.io/your-senior-devs-dont-scale-your-opencode-agents-can-e2ecf2d04548 and found it eliminates most of the "technically correct but doesn't fit" output that makes people distrust agent-written code.
Basically your prediction might already have a partial answer in the form of agents-as-code, where the domain knowledge lives in config files that travel with the repo. Did you end up seeing anything along those lines since?
We've tried variants of that (for Claude Code) and it feels hard to get right. Right, meaning, CC will reliably follow this advice and not just ignore it half the time. I think maybe there's a skill issue in how we are writing the instructions.
But I'm also curious how this works outside of coding, in fields that are less defined by plain text files.
I noticed lately that explicit skill invocation works better when it's frontloaded, especially with sonnet models