Jack Humphrey

The org chart said 40 engineers. The roadmap got 12. The other 28 were keeping the lights on. More than once in my career, I've been handed some version of the same mandate: help us improve engineering velocity, in partnership with engineering and design. The framing always assumed velocity was a process problem. Better prioritization. Tighter planning. A few weeks of looking turned up the same answer every time: years of accumulated systems, multiple pivots, and a maintenance burden that crept up year after year. Process was not the major constraint. Capacity was. Here's why this matters now. Everyone is pointing AI agents at new feature work. That amplifies the 12 and does nothing for the 28. And the scarce input in the agent era isn't agents. It's engineers. 12 engineers directing 100 agents each is 1,200 agents of supervised work. 40 engineers directing the same 100 is 4,000. I wrote about how to run this play: how to measure your deployable fraction, where the ripest agent targets hide (runbooks that are documents, the systems only one person understands), and why the agents should go to the KTLO team rather than an AI tiger team. Honest question for the engineering and product leaders here: do you know your deployable fraction? Many leaders I've asked have never truly computed it. https://lnkd.in/gQHbp-iC

17

Across four posts: no switch, no readable regular-versus-irregular category, no dedicated component, and no rule direction that even a gradient-descent optimizer could find. When the strongest available search for a thing turns up only tricks that game the metric, at some point that's an answer rather than a failed search.

In program reviews, engineers always hit technical debt first because outdated interfaces paired with crusty legacy code create headaches when bolting on new tools. But something deeper holds everybody back. Let's call it innovation debt. You see it when leaders sometimes fail to help engineers adapt to their systems, and the outfit can slow to a crawl. That is why those same legacy systems take fire in briefings, even though they have decades of battle-tested logic baked in. Engineers hammered them out under insane deadline pressure, where one bug could mean lives lost. No wonder teams rush to swap them out and lose all that tribal knowledge in the process. Programs typically crater from knotted-up ops, regs, and burned-out folks carrying gaps nobody wrote down. Look at the fallout, and you see complexity running too many programs. True change sticks when leaders respect engineering systems and culture by training them for tomorrow's challenges, even though we still tend to measure technical debt mainly by lines of code. Meanwhile, innovation debt usually shows up as burned-out engineers and sluggish processes that can break trust. The actual fix means peeling back those layers instead of chasing shiny quick fixes.

15

A coding agent can stay within its permissions and still expose data if nobody reviews what it does. While working as a fractional Director of Engineering at an AI safety company, I saw customer data sent directly to an LLM for analysis. The system had the right access, though the data flow was not reviewed. With full repo access, an agent can also reach stored secrets and credentials. We added customer-specific redaction policies to remove sensitive data before it reached the model. Agent-written code still needs human review, especially when it handles customer data. #AIAgents #AISecurity #AIEngineering #AgenticAI #SoftwareEngineering

1

1 Comment