Madness, genius, dealing with grief, or just a way to get more done
Madness, genius, dealing with grief, or just making the most use of the time we have.
In a past job, I worked for a company which traditionally would blend labor rate with offshore. However, the clients required all workers to be US Citizens or Green Card holders, so they would hire smart and capable individuals straight out of a coding boot camp but who may not have been exposed to race conditions, etc. I was pushed to see how many projects I could do at once (3 clients, so 3 parallel workstreams). Mentoring these individuals (and the learnings go both ways), building rapport and then trust, so that I could do what was three 40-hour jobs in a normal day.
That’s kind of what working with agents feels like this year. I’ve had a long career and have had the pleasure of working with some truly talented developers. A Fable PR review is impressive and my private lab go-to had been Opus orchestrator, Sonnet implementer (I’ve tried plenty of OSS tools, but Sonnet wasn’t a primary usage drainer) and Fable for PR reviews. Then again, I wanted to see how fast I could go. I posted last week about a fleet process (https://lnkd.in/p/gjzhY4YQ). At a high level you work on a plan for a set of features which may touch different repos, checking to make sure it doesn’t conflict with any past ADRs, and generating a flow spec if the API will change. Most of this is built, except for the part where the tester (a different model family, so starting with Gemini and ChatGPT) sees only the flow spec and not the actual code. I did my due diligence last year and have pretty much let Claude within the inner sanctum. I’m not willing to let other vendors in that closely because it doesn’t make sense.
But agents come with a number, you can track through vendor (or through fleet). People have a number too, but we generally don’t think in those terms. So I’m gathering metrics right now but leaning to Fable for reviews of flow-specs, ADRs, and plans rather than pull requests (relying on independent testing of the flow spec in the form of regression tests for final machine decision). I’m curious what other people are seeing, and I’ve decided to open source the fleet, but not the judgment that led to it (basically what built the trust and rapport).
The code is at github.com/botelle/eunomia (Apache-2.0), with a short description on the Free Software page. The design behind it: Merge is ignition.