if the agent environment has a bash (like your service), my take is CLI is more effective. for server environments like mentioned in the other comment, I think MCP has a place.

Providing ways to empower the agent confirm a hypothesis has been very helpful, would it be just a "make test". Then you can prompt it to test ideas and it can get a true/false signal. For a wifi driver I assume this is not accessible.

yeah it may require a few iterations. but once it gets good, then it's really possible to make more deep improvements. It's also useful to give the agent with "resources" to better understand how everything fits together, what are the unstated assumptions, etc.

Having a docs/REQUIREMENTS.md and docs/PLAN.md and asking it to keep them up to date has tremendously helped. They are referred to in AGENTS.md. It's really good at keeping them organized. I generally one explicit clean up session every 10 changes or so.

upon seeing this exchange, my daughter asked me "what age do you have?"

it's been there for a while before I left and I left in 2023

there was a big internal uproar when the team had decided to remove it (in 2022 or so?). I believe the removal didn't even have time to get to external production

Since I unsubscribed, I'm now getting in the low tens of views. Gotta love talking to myself.

The fascinating part is that posting a paying user was super flaky and slow, while as a free user it's fast. I suspect they just send the free user posts directly to /dev/null.

for your product my guess would be instagram

The amount of work that has been done to make git-on-borg scalable has been incredible. In the end, it was better for the chromium project to pre-compute the pack file every other commit and serve this file as-is over GCS and have clients fetch this all the time.

Paying for marketing? it's a good "deal" if you are targeting a specific crowd.
The feed is overwhelmed by these people. Replies even more so.

I have bad news for you.

The next step is to provide an environment where debugging can be a reasonably fast feedback loop. The initial obvious thing is "fix the build". The next less obvious one is "fix the integration test". Keeping them fast is left as an exercise to the reader. :)