“I wish I could see the prompts that led to changes”
One problem with unleashing the agents on Github is that you’re going to get a lot of trash PRs that swamp open-source maintainers. To that end Mitchell Hashimoto (of Hashicorp) has Vouch, a system for maintaining a list of authorized contributors. His point is that in the past, the effort of understanding a codebase and submitting changes was high enough that we didn’t need this, but with agents pushing code this is no longer the case. His idea is to make trust explicit.
But another problem with not just Github but git itself in an era of agentic development is that the whole model of commit messages is probably insufficient to track what’s happening. Armin Ronacher points out that it would be nice to see the prompts leading to changes. We’re now programming non-deterministic assistants to do work for us, and having a prompt trail can help us improve our own prompting workflows, as well as track where agents go off the rails.
Personally I’m tracking everything in a JSON task list autogenerated from a markdown PRD file, and this works ok for a more-or-less sequential, iterative development loop with one agent at a time. I’ve also written about Beads, which is Steve Yegge’s more battle-tested tracking system. Tying solutions like these to commits, as well as keeping record of user requests and agent responses, which can in turn be scanned by future agents to infer more about a codebase, user intent, bugs etc seems like more than git itself can handle. I’m curious to see what the software development converges on because we’re all working on larger and more complex projects by the day.