Agent HQ and the Parallel Agent Model

▶ Watch (03:31)

GitHub has been running coding agents internally for over 18 months. The conclusion: “Bolted on Coding agents simply don’t work.” Agent HQ is the answer. It is a Mission Control dashboard where developers assign tasks, track progress, and steer agents across every project in parallel, switching between CLI, web, and IDE without losing context.

GitHub’s secret scanning team shipped validators for 90 new providers in one month using Copilot Coding agent, work that would have taken quarters otherwise. In five months, Copilot-created pull requests crossed 1 million merges in GitHub’s own codebase.

Copilot Coding Agent in Action

▶ Watch (09:01)

Jessica Deen walked through the Agent HQ Mission Control demo (10:14) live on stage. She kicked off a mobile wallpaper session from the repo, opened Mission Control to see every active agent at once, and steered one mid-run with updated image dimensions. A custom Docs agent, configured by the organization, handled documentation in parallel.

The session logs showed Copilot iterating internally: building, running CodeQL, checking linters, and running secret protection before producing a PR. The PR included a full description plus Playwright-generated screenshots. Moving to VS Code took two clicks.

Third-Party Agents Join the Platform

▶ Watch (15:09)

Agent HQ is not a GitHub-only fleet. Anthropic, OpenAI, Google, Cognition, and xAI are bringing their agents onto the platform by end of year. Each gets a real GitHub identity, runs on GitHub Actions compute, and sits under enterprise audit and access controls. Mike Krieger showed Claude picking up issues, creating branches, and responding to PRs through the Claude Agent SDK, with skills loaded dynamically based on task context.

Alexander Embiricos ran the Codex parallel agents demo (22:18) in VS Code Insiders: four Codex agents working simultaneously on a Three.js game, available today for Copilot Pro+ subscribers.

Plan Mode and Custom Agents in VS Code

▶ Watch (29:55)

Plan Mode turns Copilot into a design partner before any code is written. Pierce Boggan ran the Plan Mode and TDD custom agent demo (31:12): Copilot analyzed the codebase, generated a step-by-step plan for a wallpaper-sharing feature, and handed it to a custom TDD agent. That agent, defined in a file checked into the repo, wrote failing tests first, implemented the feature, then confirmed all tests passed before surfacing a PR.

Custom agent files define prompt, tool scope, and preferred model. Every developer on the team inherits the same agent.

Agentic Code Review and Enterprise Controls

▶ Watch (38:19)

GitHub merged over 43 million pull requests per month last year. Kyle Daigle ran the agentic code review live demo (39:45): next-generation Copilot Code Review pairs model reasoning with CodeQL, flags security issues, and opens a new PR with fixes applied. No chasing down reviewers on vacation. Enterprise admins can enable it today in preview features.

Three enterprise controls shipped alongside it: a Copilot Metrics dashboard in public preview, a GitHub Code Quality view tracking maintainability inside PRs, and a new AI controls plane with audit logs, agent access scoping, and a custom admin delegation role.

Q&A

What excites you most about AI changing software development, and what do you see for developers and agents in a few years? Nadella framed the current shift as a third toolchain era: assembly gave way to compilers, compilers to code generation, and now agents generate code while developers still think in code, making tooling quality the deciding factor. ▶ 48:05

Agent HQ makes GitHub the home for agents too. Why is that shift important, and how do you see GitHub’s role going forward? Nadella said the Agent HQ metaphor threads Git, PRs, Issues, and Actions into a workflow only GitHub can own end-to-end, and that without quality tooling, exponential model improvement still yields sublinear productivity gains. ▶ 51:01

What skills will be useful for this new generation of developers? Nadella said the ‘what’ of desired outcomes has not changed, but the ‘how’ has shifted entirely, and developers need to build metacognition around the new toolchain the same way he once had to learn to think with Lex and Yacc. ▶ 56:36

Notable Quotes

Bolted on Coding agents simply don’t work. Kyle Daigle · ▶ 4:57

The GitHub workflow is the agent workflow. Kyle Daigle · ▶ 14:56

Copilot is already the number one contributor to real merged pull requests Kyle Daigle · ▶ 6:55

a new toolchain is being born right in front of us. Satya Nadella · ▶ 50:13

Key Takeaways

  • Agent HQ lets developers run parallel Copilot sessions from one dashboard, steering agents mid-task without switching tools.
  • Claude, Codex, and other frontier agents get a real GitHub identity and run on GitHub Actions compute inside Agent HQ.
  • Agentic code review combines model reasoning with CodeQL to catch security bugs before they land, and can apply fixes automatically.