Google engineer: “In 2026, if you aren’t building AI agents, it’s crazy how far behind you already are. At Google, 85% of our engineers were running self-improving agent harnesses.” In this 32-minute talk, a Google engineer with 30 years of experience explains what the future of agentic coding will... look like. Worth more than 10 paid agent engineering courses. Watch it today, then read the guide below to build your first production-grade agent harness ↓show more

The biggest shift in AI coding is moving from writing code faster to building systems that can reliably produce and improve code.

How many things to watch man .. I see every random post .. watch it today .. else you’re missing it .. enough already

Harness design will age better than the 2026 prediction.

"Self-improving agent harnesses" is the specific phrase that separates a loop that retries from a system that gets better — and that distinction is worth 32 minutes of anyone's time.

I play this game where I open 6 terminals in different terminals in different accounts across 2 Mac’s and I don’t warn any account about the other. Flawless execution every time

The real shift is that coding skill is becoming less about writing every line yourself and more about designing reliable agent harnesses that can plan, execute, verify, and improve.

85% already doing this, wow

This interview is old and it still hits home

harness over prompt is right. but 85% counts engineers who ran one, not work that shipped out of one. and that second number never travels with the first

Branch conditions in my runner parse to exactly true or false. A model that writes 'probably yes' fails the step instead of quietly picking a path. That catches the malformed answer, not the confidently wrong one. Smallest layer I added and the loudest.

The shift from AI that assists to AI that acts is happening fast. Learning how to build and manage reliable agents could become one of the most valuable skills in software engineering.

85% is a crazy number tho

Building AI agents now is no longer optional for serious engineers

The durable shift is from prompting a model to operating a controlled loop: scoped context, tools, tests, and recovery. If an agent cannot show its evidence and stop safely, more autonomy just scales the failure rate.

sounds impressive but i've seen more demo than deployment

Getting an agent into production is one challenge. Operating it safely at scale is another — especially when it has access to real tools, data and APIs. #AIAgents #AIsecurity #AIAgentControl

85% of google engineers can't even agree on a linter

this is a wake up call for anyone still coding without agents!

A smart agent without a good harness is still a fragile system.

Self-improving harnesses are useful only if improvement cannot silently expand authority. Changes to prompts, memory, tools or routing should be versioned, evaluated against failure cases and promoted through an approval boundary. Otherwise the system learns operationally while governance remains static.

everyone is still comparing models while the best engineers are building better harnesses around them

Just create a project. So being so lazy people