Vice Software, LLC ads

Two days in Chicago with a bunch of other engineering leaders, and the conversation was 180 degrees from where it was 12 months ago. A year ago this room would have spent a lot of time on whether AI can write production code at all. Not anymore. This time it was all downstream of that. What software looks like when you stop optimizing it for humans to read. Whether speed versus security is even a real tradeoff, or whether the actual choice is governed velocity or unmanaged velocity. How much of the quality problem you're chasing at the end is really a spec problem at the beginning. Whether QA as a function survives any of this. I talked about the three phases of agentic engineering. What phase your org is actually in, what breaks on the way up, and what your developers should be doing instead of babysitting agents. The thing is, four talks from four different angles and they all landed in the same place. The tooling stopped being the hard part. The org is the hard part now. Thanks to 7CTOs for putting it on, JDAQA for sponsoring, and Ben F., @Tyler Ward, Greg Svitak and Jay Aigner.

This Friday! (for real this time) The 3 phases of agentic engineering. Where most teams actually sit, and what breaks on the way up. Driver, Supervisor, Director, and the three-agent ceiling in the middle that stalls most teams. LinkedIn Live, Friday July 17 at 2 PM CT. Live, not recorded. Bring your questions. Register here: https://lnkd.in/gNBKDc6E

I built my agent so it can't stop and ask me dumb questions. Claude has a stop hook. When the agent wants to stop and check in with you, you can override it. Intercept it. So the agent comes to me wanting to hand off a PR. Before it reaches me, the harness asks it. Did you run the tests? Did you run the type checking? Did you follow the patterns? If the answer is no, the harness says "sorry, Ryan doesn't want to talk to you right now. Get back to work" The agent only reaches me when the work is actually ready. And if it gets genuinely stuck, I don't want it burning tokens forever either. So we track it. Same failed approach three times in a row, the agent pauses itself and posts in Slack that it's stuck. Then a human steps in. That's harness engineering. A policy layer that makes the agent do the things you'd otherwise have to nag it about. Most teams are still doing the nagging by hand. I'm walking through my thoughts on a LinkedIn Live this Friday, July 10 at 2 PM CT. The three phases, the 8 levels of maturity, and how the harness gets you past the three-agent ceiling. Bring your questions. Register here: https://lnkd.in/dBGpEp8u

BOOM... New level unlocked. 10 agents running 10 tickets in parrallel! This is on a new client project where they were looking to level up their team. We are 4 days in and took them from 1 agent "driver" flows to 10 agent flows with my custom harness code. Fortunately I met with QA and Product this week to help optimize their flows because dev velocity is about to go through the roof and if we don't get ticket velocity and testing velocity up then we just moved the bottle neck. Never a dull moment!

I think there are basically three phases of agentic engineering. Almost nobody gets past the second one. Everybody starts in the same place, and the progression is real. Bassim Eledath wrote it up as eight levels in a piece that made the rounds, worth reading. But when I watch teams actually move through it, eight rungs is too many. What changes is your role. And your role only changes twice. Phase one. Driver. Tab complete. Copilot. Then a real agent, Claude Code or Codex. Then you hit the dumb zone, where the more you ask the agent the dumber it gets, so you learn to manage context. That's the early stretch. All of it is you driving one thing at a time, focused the whole time. One car, your hands on the wheel. Phase two. Supervisor. You start writing things down. Specs. Documentation the next agent can pick up. I was doing this two years ago and calling it documentation-driven development. You pull in MCPs and skills. Now you're running three agents at once. This is where almost everyone stops. Three agents, and you spend the whole day micromanaging. The supervision tax eats the gain. Every question routes to you. I maxed out here. I haven't watched anyone doing these approaches get past three either. The closest I ever saw was a guy running four. He had an eye-tracking rig wired up so he could flip between terminal windows just by looking at them. Wildly impressive. Also kind of insane. Short of turning yourself into a comic book cyborg, three is the wall. Here's the part people don't want to hear. Everyone's chasing background agents. Work that ships while you sleep. You cannot get there as a supervisor. More agents doesn't break the ceiling. It just spreads you thinner. Phase three. Director. This is harness engineering. You stop supervising and build the supervision into the harness. The type checks, the tests, the done criteria. The harness fields the questions. You review the plan and check the work. The agents don't come back until they pass. A supervisor watches every agent. A director sets the constraints and lets the harness do the watching. That's the unlock. It's the only door to the autonomous stuff everyone wants. Three phases. Driver, supervisor, director. The whole game is getting out of the supervisor seat. What are you trying to add right now, more agents or less supervision?

Doing a LinkedIn Live this Friday at 2 PM CT. Walking through the 8 levels of agentic engineering maturity. Where most teams actually are, where they think they are, and what's keeping them stuck at the three-agent ceiling. If you've been following anything I've posted about harness engineering, this is the full picture in one session. Live, not recorded. Real questions, not a slide deck with pre-written answers. See you Friday. https://lnkd.in/g-DnUj9x