← Google
Cole Medin
Google Just Dropped a Masterclass on Agentic Engineering (It's SO Good)
⏱ 22 min video · 4 min read30 Jun 2026Worth watching
TL;DR
Cole Medin breaks down a 51-page Google masterclass on AI-driven software development, covering the spectrum from vibe coding to agentic engineering. The core argument is that the 'harness' — your rules, workflows, tools, and guardrails — matters 10x more than which LLM you pick, and building it well is the highest-leverage investment an engineer or team can make.
Key points
1
The AI-driven SDLC shifts the bottleneck from implementation (now minutes/hours with AI) to requirements gathering and validation — those stages are not significantly faster yet.
2
AI coding exists on a spectrum: vibe coding, structured AI-assisted, and agentic engineering — you pick the right level for the job, not always the most advanced.
3
The harness (instructions, MCP servers, guardrails, hooks, test infrastructure, observability) accounts for 90% of system quality; the LLM model itself is only 10%.
4
Static context (system prompt, core rules) should be kept lean and loaded every session; dynamic context (agent skills, workflow docs) is loaded on demand to avoid context rot and reduce cost.
5
Vibe coding has low upfront cost but high operational token burn; agentic engineering requires upfront investment but becomes 3-10x more reliable and cheaper over time.
6
You only need one generalist agent — agent skills via dynamic context replace complex multi-agent specialist architectures.
Actionable insights
→
Build your harness before scaling AI coding: invest time upfront in specs, rules, workflows, and CI gates — the crossover point where it pays off arrives very quickly.
→
Keep your system prompt lean; offload specialized knowledge (code review, planning, conventions) into loadable agent skills so the context window stays clean and costs stay low.
→
Split every coding task into two separate agent sessions — a planning agent to create the spec artifact, then a coding agent to execute — to prevent context rot and bias bleeding between phases.
→
After every iteration where the agent struggles or you have to intervene, ask the agent to retrospect and suggest improvements to the harness rules or workflows (system evolution mindset).
→
Use terminal bench 2.0 benchmarks as evidence when justifying harness investment: LangChain gained 13.7 points (the gap between Sonnet and Opus) purely through harness engineering, not model switching.
Notable quotes
“The large language model that you use for your AI coding assistant is only 10% of the system — everything else like your instructions and tools and context and guardrails and orchestration and observability makes up the other 90%.”
“You can make Sonnet work as well as Opus if you have the right system, the right process that you are having it go through as the harness.”
“The specification quality is the new bottleneck.”
Worth watching?
✅
Worth watching the full video?
Watch if you want the visual diagram walkthrough and Cole's commentary on where he disagrees with Google — the key frameworks and numbers are all captured here, so the summary alone may be enough for most readers.
Saved you some time? The creator still deserves a like.
Want this for your own channels?
Add the channels you follow. Every new video summarised and in your inbox the moment it drops. From £4/month.
Try it free