AI-Wiki

Agentic DevOps in Real Life – Build Faster, Ship Safer, Keep Humans in the Loop

7 min read

Agentic DevOps in Real Life – Build Faster, Ship Safer, Keep Humans in the Loop

In this Live! 360 keynote, Brian Randell and Mickey Gousset unpack how agentic AI is reshaping DevOps—from coding and code review to automation, security, and incident response. You’ll see how GitHub Copilot, agent workflows, and new Azure capabilities let teams automate the repetitive work, keep humans in the loop where it matters, and accelerate delivery without sacrificing safety.

Through real demos in GitHub, Azure DevOps, Visual Studio Code, and Azure’s new “agent” preview, you’ll learn how agents collaborate with developers and IT pros, where they excel, where they fail, and how to build a practical 30/60/90-day adoption plan for rolling this out in your own organization without losing control, quality, or security. — Microsoft Visual Studio channel description

A ~58-minute Live! 360 (Orlando 2025) keynote on the Microsoft Visual Studio YouTube channel by Brian Randell (partner at MCW Technologies; ex-GitHub; long-time Microsoft MVP; co-author of Essential DevOps) and Mickey Gousset (Staff DevOps Architect at GitHub). A Microsoft/GitHub vendor keynote, demo-heavy, pitching “agentic DevOps” — the integration of AI coding agents into the full software-delivery lifecycle while keeping humans in the loop. The wiki’s first dedicated GitHub-Copilot / Agent-HQ tooling source and the trigger for promoting GitHub and Microsoft to entities.

TL;DR

  • “Agentic DevOps” defined. “The next evolution of DevOps reimagined for a world where intelligent agents collaborate with you… to optimize and automate every stage of the software life cycle.” DevOps is framed as “the union of people, process, and products” for continuous delivery of value — and “an organizational transformation first.” The load-bearing word is collaborate: agents extend the team (interactive and autonomous), they don’t replace it.
  • GitHub Copilot coding agent + Agent HQ. Assign a GitHub issue (or Azure DevOps work item) to Copilot → it opens a draft PR, works in the background (reads files, runs tests/bash), and returns for review. Agent HQ (announced at GitHub Universe) is the centralized surface to monitor, steer an active session mid-run, and audit past runs. The open Agent HQ architecture lets you run the OpenAI Codex agent and (coming) Anthropic Claude / Gemini agents from inside VS Code without a separate vendor account — included in the Copilot license.
  • VS Code four modes: Ask / Edit / Agent / Plan. Plan mode is the newest — the agent states how it will implement a task before acting, to bound the “too much autonomy” failure mode. Instructions files encode house rules + exemplars; agents run locally (CPU/GPU) or get pushed to GitHub’s cloud. Offline/local agents via Ollama + the Continue extension.
  • Responsible-AI pipeline + human-in-the-loop. Copilot routes requests through Microsoft’s responsible-AI pipeline, validating both input and the returned code (security vulnerabilities, NSFW) — even when the chosen model is third-party (Anthropic/Gemini), in milliseconds. The hard safety rule: by default the person who initiated a Copilot request cannot be the final approver to merge into the parent branch. “AI suggests, humans decide.”
  • Security. GitHub Advanced Security = secret scanning + code scanning, now with Copilot autofix (“found means fixed” — a generate-fix button). Stat: GitHub secret protection blocked 4.4 million credential leaks in 2024. Randell’s vibe-coding anecdote: Claude wrote a browser extension, Codex reviewed it, both passed it — and Advanced Security still caught a URL-injection vulnerability multiple times. “Two smart AI agents still screwed up and GitHub had my back.”
  • The productivity paradox (the METR study). “We have the expectation that AI is going to help us go faster. But sometimes you have to go slower to go faster.” Developers measure slower on average while feeling faster — the “paper-pushing” illusion of busyness (more files spinning out ≠ more shipped value). Autobahn analogy: speed needs skill, risk tolerance, and knowing the rules. There’s a trust gap to close by working with the tool in your own domain. “AI is powerful but not magic.” Corollary: don’t chase every new frontier model — each switch costs learning time.
  • Jobs. Displacement is real (Randell’s 1980s typing-pool analogy) but framed against the World Economic Forum’s 78 million net new jobs by 2030 projection — the point is collaborative use, not replacement.
  • AI-code risk stats (presenter-cited, attribution promised in the deck): “322% more privilege-escalation paths in AI code,” “40% increase in secrets-exposure risk,” and “AI-assisted commits merged 4× faster, bypassing reviews.” Frontier models are trained on the whole internet — including insecure code; agents tend to offer the cheap/easy auth (passwords/PATs) over safer OAuth.
  • 30/60/90-day adoption playbook. Progressive ROI, don’t boil the ocean, break the work down for the agent like you would for a human (giving it too much at once is the common failure). Role-by-role upskilling (developers: prompt-writing, code-review, security; DevOps/security teams; get the C-suite to upskill too). Agentic patterns: interrupt-and-resume, human-as-a-tool, approval-flow, fallback-escalation. Collaboration spectrum: human-in-the-loop / on-the-loop / out-of-the-loop (the presenters use nothing fully out-of-the-loop).
  • Azure SRE Agent (preview). An autonomous cloud-monitoring + incident-response agent: scope it to a resource group, it does incident management, can be permissioned to autonomously fix certain conditions, and files health/security reports — “a site reliability engineer on my team.”
  • Five takeaways: start Monday with an action plan; security first; keep the human in the loop (AI suggests, humans decide); measure everything; focus on value, not lines of code generated.

What was actually ingested

The full ~58-minute auto-captioned transcript (the caption track looped on fetch — trimmed to the unique 0:00–57:43 content; ASR-cleaned, see the raw file’s notes:). As a vendor keynote, the product capabilities (Agent HQ steering, open-agent architecture, Azure SRE Agent, autofix) are presenter demos and claims; the cited statistics (4.4M leaks blocked, 322% privilege-escalation, 40% secrets-exposure, 30% acceptance / 84% successful-build) are stated verbally with attribution deferred to a conference PDF — treat them as presenter-reported. The slide deck itself was not captured.

Dynamic-capabilities reading

  • digital-transforming/redesigning-internal-structures — the core thesis: agentic DevOps reshapes the software-delivery workflow and team roles (issue→agent→PR→human-review), with Randell explicit that “DevOps is an organizational transformation first” and “everyone’s responsibility.”
  • digital-transforming/improving-digital-maturity — the 30/60/90-day playbook, role-by-role upskilling, and the “measure everything / progressive ROI” framing are a maturity-ramp prescription for adopting agentic workflows across an organisation.
  • digital-seizing/rapid-prototyping — the Copilot-coding-agent and VS Code Agent/Plan demos (assign an issue, get a tested PR; vibe-coding a browser extension) are rapid build-and-iterate loops; the “spin it up in the cloud or run locally” optionality is prototyping agility.
  • strategic-renewal/organizational-culture — “keep the human in the loop,” the initiator-can’t-self-approve rule, and the “AI suggests, humans decide” / “focus on value not lines of code” maxims are a cultural prescription for how teams should relate to agents (trust built through use, not blind acceptance).
  • contextual/internal-barriers — the productivity paradox (feels-faster-but-slower), the trust gap, DBA/IT-pro resistance to automated deployment, and the AI-code security-risk stats are named adoption barriers the keynote warns against.

Linked entities and concepts

  • GitHub — promoted to an entity on this ingest (Copilot, Agent HQ, Advanced Security, the State of the Octoverse report; Gousset is GitHub staff). “GitHub is the center of the universe for Microsoft’s developer agentic tooling.”
  • Microsoft — promoted to an entity on this ingest (channel publisher; Azure, Azure DevOps, VS Code, the responsible-AI pipeline, Azure SRE Agent).
  • OpenAI — the Codex agent, now runnable inside VS Code via open Agent HQ.
  • Anthropic — Claude Sonnet 4.5 named as a favoured frontier model; Claude agent coming to Agent HQ.
  • Google — Gemini named as a selectable model / coming agent.
  • METR — the productivity-paradox study anchoring the “feels faster, measures slower” claim.
  • Concepts updated: micro-productivity-trap (the productivity paradox + “value not lines of code”), responsible-ai (the responsible-AI pipeline + human-in-the-loop merge rule + AI-code risk stats), agentic-engineering (Agent HQ, the four VS Code modes, plan mode, autonomous agents), enterprise-ai-adoption (30/60/90 playbook, trust gap, upskilling), vibe-coding (the vibe-code→Advanced-Security-caught-injection anecdote), ai-employment-effects (displacement vs the WEF 78M-jobs projection).
  • Dangling (single-source mention, deferred per author-entity rule): Brian Randell (presenter; MCW Technologies; Essential DevOps), Mickey Gousset (presenter; GitHub Staff DevOps Architect), GitHub Copilot (product), Agent HQ (product), Azure SRE Agent (product, preview), Azure DevOps, Visual Studio Code, Live! 360 / VS Live (venue), World Economic Forum (cited for the 78M-jobs stat).

Relationships

Supports

  • AI-Native Software Engineering: Enduring Principles, New Pace — Both argue agentic AI changes the *pace* of software engineering while the *enduring principles* (review, security, human judgment) still bind — and both cite the METR productivity finding that AI-assisted developers feel faster but measure slower. Randell/Gousset is the vendor-practitioner-keynote altitude (GitHub Copilot + Agent HQ tooling demos); Sinclair/Ivers/Benitez is the institutional-research-centre altitude (SEI-CMU). Convergent on *keep-the-human-in-the-loop / measure-don't-assume*.
  • AI Dev 26 x SF | Anush Elangovan: Impact of AI on Software — Same agentic-software-engineering-adoption phenomenon — agents writing/refactoring code at speed, the human role shifting to review and orchestration. Both reference the METR productivity-paradox evidence. Elangovan is the AI-infra-vendor (AMD/ROCm) altitude; Randell/Gousset is the developer-tooling-vendor (GitHub/Microsoft) altitude.
  • Running an AI-native engineering org — Both land the *the tool is not the hard part — your processes and trust are* thesis for org-wide agentic-coding adoption. Anthropic's *Running an AI-native engineering org* is the model-vendor-customer altitude; Randell/Gousset is the Microsoft/GitHub-tooling altitude with an explicit 30/60/90-day rollout playbook.

Graph View

Backlinks