Agentic Coding, the Boring Way

TL;DR

• We're in an "agentic hangover": citing Gartner research, using more AI doesn't produce more productivity — it mostly makes Anthropic and OpenAI richer.
• Two myths to kill: (1) agent swarms — hundreds of agents on one problem, yielding an explosion of both cost and errors; (2) loop engineering / loops within loops — same bill, same outcome.
• Cheap models won't bail you out: another study on AI code generation shows uncontrolled use leaves more errors, and more persistent errors, in production.
• The real hidden costs are verification, security holes, debugging, and finally cognitive debt and burnout — lean on AI too hard and you lose the context.
• The fix is the D3 framework (built at Amazon in 2025): Discover (a research artifact covering the repo plus upstream and downstream dependencies) → plan (well-defined, broken into subtasks) → Deliver (AI executes). Plus human sign-off, multiple models from different companies rather than one family, and no more than about four well-scoped agents.
• One takeaway: stop token maxing, start ROI maxing.

Key Points

Why legacy, and why "boring" (~00:48–00:49)

He scoped the talk up front: this is about legacy systems in enterprises. The "boring way" is for people who want systems that are scalable, secure, and reliable.

His background: former head of AI at Amazon Germany, where he built large-scale AI systems worth billions in revenue impact, plus work in robotics AI. He now runs Omokai, building a physical AI operating system that turns voice into autonomous missions for robots and drones. He wasn't there to pitch Omokai but to describe how his engineering team works — and warned that some of it would be controversial.

A telling detail about that team: Omokai is deeply AI-native — custom fine-tuned models, AI-generated data, and seven human employees against more than ten "AI employees."

The diagnosis: an agentic hangover (~00:50–00:52)

• The productivity paradox: citing Gartner research, more AI does not translate into more productivity — it translates into Anthropic and OpenAI getting richer. If you're using AI unbounded, without proper constraints, you're probably using it wrong.
• Two fashionable myths:
• Agent swarms — hundreds of agents on a single problem, which buys you a lot of cost and a lot of errors.
• Loop engineering / loops within loops — he's seen people nest loops, and it costs far more than you'd want to pay.
• "AI will get cheap, so cost won't matter" doesn't survive either: another study on AI code generation found that uncontrolled use produces far more errors — and more persistent errors — in production databases.
• The hidden cost chain: verification, then security loopholes, then debugging, and finally cognitive debt and burnout. He insists this last one is real: use AI too much and you end up losing a lot of context yourself.
• He showed a screenshot of someone screaming at the internet, fed up with an open-source repository drowning in AI slop. "We don't want to do that."

The prescription: the D3 framework (~00:52–00:53)

What he learned at Amazon is that frameworks and mechanisms are what make systems reliable and scalable. In 2025 he built a D3 framework for agentic coding at Amazon, with three phases:

1. Discover — produce a research artifact containing everything about your repository, plus its upstream and downstream dependencies.
2. Plan — from that research document, write a very well-defined plan broken into subtasks.
3. Deliver — the AI takes all of that information and solves the problem.

Three supporting rules:

• Don't fully trust AI: keep a human sign-off in the loop.
• Diversify across vendors: use multiple models from different companies, not variants from the same family.
• Don't use too many agents: around four, each well defined.

Closing: if you take one thing from the session — stop token maxing and start ROI maxing.

Quotes

"Using more AI does not lead to more productivity. It just leads to you making Anthropic and OpenAI richer." (~00:50)

His blunt summary of the Gartner finding.

"If you use AI too much you end up losing a lot of context." (~00:51)

Cognitive debt in a sentence: it's the human who loses the context, not the model.

"Stop token maxing and start ROI maxing." (~00:53)

The one line he asked the audience to take home.