I’ve written a few of these now: the AI working charter, the skill atrophy piece, the security post, and the deep dive on the agent orchestration framework my team built. This one’s different, because it’s not about a framework or a policy, it’s about something I realized I got wrong in how I was leading through all of it.
I was listening to Kent Beck on the Pragmatic Engineer podcast (the Kent Beck of TDD, Extreme Programming, one of the founding names on the Agile Manifesto), and he laid out a framework he calls explore, expand, extract. Explore means you don’t know what’s going to work, so you run a lot of cheap, uncorrelated experiments. Expand means something clicks, and you drop everything else to focus on it, and extract means you’ve got a proven and repeatable playbook, so now it’s about efficiency at scale.
His point that stuck with me was that for the last 20 years, software engineering has mostly lived in extract mode, since there was a known playbook and being a senior engineer largely meant knowing it. Agentic coding tools wiped that playbook clean, and nobody has the new one yet, not us, not Anthropic, not whoever’s loudest about it on LinkedIn this week (myself included). Beck’s own reference point is that the Agile Manifesto took 15 years of doing object-oriented programming before anyone could distill it that simply. Depending on where you draw the line, whether that’s Devin making “agentic coding” a term people actually used, or Claude Code and MCP getting good enough to actually trust, the industry itself is somewhere between one and two and a half years into this. My team is six months into doing this at scale as an org, even if a few of us were tinkering with it longer on our own. Even the timeline is unsettled, which is kind of the point.
And this is where I had to check myself. My team is at wildly different stages of this shift, with some people still mostly using autocomplete and chat while a few of us went deep and built the agent orchestration framework I posted about a few weeks back. I realized I’d been holding a model in my head, something close to Steve Yegge’s 8-stage developer-agent evolution, and quietly mapping my team onto it like the AOF group was further along and everyone else needed to catch up. Worse, I’d been wielding the framework as “the way,” intentionally or not.
Entirely the wrong frame. We’re all in explore mode right now, because whether you’re on autocomplete or running a five-agent pipeline, those are different experiments running in parallel, not different points on a ladder (Yegge’s or anyone else’s). If I pushed everyone to converge on the framework we built, we’d learn less over the next few months, not more, since we’d lose the variance that tells us what’s working.
So this week I told the team exactly that. The framework itself is one bet, not the answer, and I’d rather people find the holes in it than treat it as settled. Then I gave them something specific: tried, result, would do differently. Three fields, one sentence if that’s all it takes. If you tried anything, even if it went nowhere, bring it to the share. “I tried this and it made things worse” needs to count as much as a win, arguably more, since that’s the harder thing to admit to and the more useful thing to hear right now.
If you’re leading a team through this same shift, that unevenness across your people probably isn’t a gap you need to close. It might be the best data you have right now.