Agentic Workflows as Executable Documentation

GitHub Next is shifting the paradigm of automation from rigid heuristics to agentic workflows. Instead of complex YAML configurations, these workflows are defined in plain English within Markdown documents. This approach treats the Markdown as the source of truth and the YAML as a compiled artifact. By using Copilot to interpret these natural language playbooks, developers can automate complex, multi-step tasks like dependency upgrades across major versions, including code changes and build verification.

Crucially, these agents operate under strict, deterministic guardrails defined in the document's front matter. Rather than relying on prompt-based instructions—which are vulnerable to injection—developers explicitly declare allowed tools, network destinations, and output types. This "defense-in-depth" strategy ensures that agents remain within a sandbox, cannot access secrets directly, and are prevented from creating excessive noise (denial-of-service) by limiting their output capabilities.

Real-Time Multiplayer Development with ACE

While current AI tools focus on the 5% of software development spent typing, GitHub's ACE prototype targets the remaining 95% of the lifecycle: planning, political alignment, infrastructure decisions, and team communication. ACE functions as a real-time, multiplayer development environment running in cloud-based microVMs.

Key features include:

  • Shared Context: The environment captures the entire history of team discussions, allowing the AI to synthesize final states from messy, iterative human conversations.
  • Collaborative Planning: Teams edit Markdown-based plans together, which the agent then executes to make the code match the agreed-upon document.
  • Ambient Awareness: The interface surfaces what teammates are working on, bridging the gap between individual coding sessions and team-wide alignment.

This shift moves the developer experience away from isolated coding toward a collaborative, document-driven process where the AI acts as a participant in the team's decision-making loop, rather than just a code-completion engine.