> The problem is that agent-written code has a provenance gap. When a human writes code, the reasoning lives in their head, in Slack threads, in PR descriptions. When an agent writes code, the reasoning evaporates the moment the context window closes.
The described situation for human-written code isn't much better. What actually works is putting a ticket (or project) number in the commit message, and making sure everything relevant gets written up and saved to that centralized repository.
And once you have that, the level of detail you'd get from saving agent chats won't add much. Maybe unless you're doing deliberate study of how to prompt more effectively (but even then, the next iteration of models is just a couple months away)?
Agent memory can sometimes be handled with prompts, state output on completion, etc. Managing that can get messy if you aren't on the same system.
I think "provenance gap" or temporal history can be helped by understanding what you have asked agentic systems to write, understand things written, and verified them.
We aren't yet at a point where something large or extended is easily pushed to agentic coding management - your point of provenance and memory is key here.
> The problem is that agent-written code has a provenance gap. When a human writes code, the reasoning lives in their head, in Slack threads, in PR descriptions. When an agent writes code, the reasoning evaporates the moment the context window closes.
The described situation for human-written code isn't much better. What actually works is putting a ticket (or project) number in the commit message, and making sure everything relevant gets written up and saved to that centralized repository.
And once you have that, the level of detail you'd get from saving agent chats won't add much. Maybe unless you're doing deliberate study of how to prompt more effectively (but even then, the next iteration of models is just a couple months away)?
Agent memory can sometimes be handled with prompts, state output on completion, etc. Managing that can get messy if you aren't on the same system.
I think "provenance gap" or temporal history can be helped by understanding what you have asked agentic systems to write, understand things written, and verified them.
We aren't yet at a point where something large or extended is easily pushed to agentic coding management - your point of provenance and memory is key here.