A Microsoft engineer's four-part guardrail kit: with guardrails you're a 20x engineer, without them you're 20x the slop.
From Writing Code to Designing Systems: How the Developer Role is Changing — Chris Noring, Microsoft · Chris Noring
23 min total·Actually worth watching closely: ~7 min·2 must-watch clips
- 0:15 – 4:00Listen
The role shift: writing code is no longer the bottleneck
The opening argument: developers move from writing all the code by hand to designing systems and orchestrating agents; the individual is no longer the delivery bottleneck, and what engineers bring management changes from PowerPoints to a running demo.
The purpose of AI is to scale developers, not to replace them.
Pure spoken argument, nothing on screen — fine to listen to on the commute.▶ Jump to 0:15Speaker · Chris Noring - 4:00 – 7:16Listen
The CLI-first workflow
Lays out a three-step workflow: start at the GitHub Copilot CLI or Claude Code, drop the editor to second place as a cross-vendor 'control board', then delegate at scale — with a look back at the lessons of the AI-slop era.
Guardrails are the price of admission for using AI tools, not an option — unconstrained output is worse than not using them at all.
Still setting up the argument; the chapter marker has nothing worth looking at.▶ Jump to 4:00Speaker · Chris Noring - 7:16 – 10:10Skim
Guardrail one: AGENTS.md
AGENTS.md is the absolute minimum guardrail, and every repo should have one: repository intent, architecture, constraints and the do's and don'ts, so agents don't drift.
You can copy AGENTS.md straight over from an old project and let the agent adapt it to the new codebase — adoption costs almost nothing.
Mostly a slide listing what the file should contain; skimming the checklist is enough.▶ Jump to 7:16Speaker · Chris Noring - 10:10 – 12:46Skim
Guardrail two: Skills as recipes
When an agent gets a repetitive task almost right but not quite, give it a Skill — a recipe and a contract that makes it follow a set order instead of improvising. The folder-plus-SKILL.md format has converged across vendors.
One line to remember as the test: almost right but not right → reach for a Skill. And if you can't word the instructions, have the AI write them.
Format and examples are on slides, no demo; skim the structure.▶ Jump to 10:10Speaker · Chris Noring - 12:50 – 15:20Listen
Guardrail three: why you need a Custom Agent
A Skill can't orchestrate itself; when you need orchestration and reasoning you move up to a Custom Agent — a higher-level concept with a specific role that can combine several Skills and plug into MCP servers.
Three steps up: AGENTS.md sets the floor → Skills set the procedure → Custom Agents do the orchestration.
A spoken build-up of the concepts; get the 'why' here, the screen comes at the next stop.▶ Jump to 12:50Speaker · Chris Noring - 15:20 – 19:51Watch
.agent files and least-privilege design
A real look at a .agent file under .github/agents: the tools field in the front matter declares what the agent may use (read, web search, building to-do lists) — a researcher agent, for example, may search but may not change files.
The tools declaration is how you enforce least privilege — what the agent can do is decided by your allowlist, not by the agent.
The first frame the LLM flagged as worth watching (938s); you have to see the screen to copy the config fields.▶ Jump to 15:20Speaker · Chris Noring - 19:51 – 22:45Watch
The delegation loop and the human-in-the-loop close
A demo of creating a GitHub issue, assigning it to the Copilot agent, letting it work in the background and getting a Draft PR back for review; the close stresses that agents swing between genius and dumb like a child, and that the human designs the approval gates and does the merge.
The real power is background parallelism — but the first three guardrail layers have to be in place before you delegate in parallel, and the merge always stays with the human.
The second frame worth watching (1198s): the full UI flow from issue assignment to work-in-progress to Draft PR — audio alone loses the shape of the workflow.▶ Jump to 19:51Speaker · Chris Noring