C-level executives with a technical past, like GitHub's COO, are using AI to build their own internal tools. This allows them to apply their unique blend of business and technical expertise to solve problems directly, bypassing traditional workflows and increasing their effectiveness.
GitHub stars and commit counts are poor trust signals because they are easily gamed by malicious actors who create aged accounts with fake activity. The solution isn't a new universal metric. Instead, maintainers should be given tools to define their own complex, agentic trust heuristics that are harder to spoof.
GitHub's recent outages are due to AI changing workload patterns. Agents create larger pushes and PRs, breaking assumptions in 15-year-old systems. This creates a 'diagonal' scaling challenge that simple vertical or horizontal scaling can't fix, necessitating rewrites of fundamental components like permissioning and job queuing.
GitHub's COO finds AI's greatest utility isn't generating new content, but performing retrospective analysis. Agents synthesize data from PRs, Slack, and meeting notes to summarize what worked and what didn't. This pattern recognition on past data is more valuable for strategic decision-making than simple content creation.
To scale AI usage beyond engineering, GitHub avoids complex new UIs. Instead, they provide a command-line interface (CLI) and shared "skills" (scripts) even to non-technical staff. This allows everyone to run powerful automations and access company context from disparate sources without changing their existing workflows.
With AI handling tasks like presentation creation, the Chief of Staff role is evolving. GitHub's COO notes he no longer needs someone to build slides. Instead, he needs that person focused on high-value human tasks: making connections, identifying opportunities, and managing relationships, which AI cannot automate.
New AI-driven code review processes haven't replaced the pull request because they miss the core point. The PR isn't just a technical workflow; it's a social protocol for codifying trust. We trust a change because a specific senior human reviewed it. Agent-driven reviews diffuse this trust, making them harder to adopt.
To get executive buy-in for AI-driven work, GitHub's COO built a revenue planning deck with an agent instructed to make it look "not pretty" and non-AI-generated. The CRO and CFO didn't notice, proving the content's value. Intentionally adding human-like imperfections can sidestep bias against AI and increase adoption.
GitHub's COO argues that future AI developer tools will be defined by their ability to pull in 'ambient' business context. Instead of just analyzing code, they need access to spec docs, emails, and conversations to make better implementation decisions. This requires a fundamental integration of developer and business information systems.
GitHub is abandoning complex, "mega-skills" for AI agents, finding large all-in-one workflows brittle and hard to maintain. The better approach is to build atomic "micro-skills"—like Lego blocks—that do one thing well. These can then be composed and orchestrated into more complex, flexible automations.
A skill like "summarize" is not one-size-fits-all. GitHub's COO notes that a summary for an analyst briefing is fundamentally different from one for a customer meeting. The key to effective AI agents is creating context-aware permutations of atomic skills that adapt to the specific audience and goal.
Microsoft's focus on open-source agents is strategic: to run agents safely at work, you need deep OS-level sandboxing. By contributing heavily to this space, Microsoft is building the foundational platform components that make Windows and Azure indispensable for the next generation of enterprise AI.