GitHub’s Agent Era: 14x Commits, 200M Developers, Copilot’s Next Act — Kyle Daigle
Summary
Agent-driven activity is simultaneously GitHub’s strongest demand signal and its largest near-term execution risk. swyx cites 1 billion commits in 2025 versus 275 million per week by April—roughly a 14-billion annual pace if growth were merely linear—and the GitHub guest says it is still accelerating. GitHub now handles “more in a month than we did in a year last year,” making this both a “hard time” and its most exciting period.
The infrastructure response requires architectural rewrites, not ordinary capacity expansion. Actions needs materially more CPU as agent-driven PR growth produces more builds; permissions remain concentrated around the database internally called MySQL 1; and larger monorepos are changing average push and PR sizes. The GitHub guest calls the new scaling vector “diagonal”—neither vertical nor horizontal—and expects “fewer and fewer” availability incidents over the next three months as step-change fixes land.
Copilot’s strategic center has shifted from code completion toward a common runtime for agents across the software lifecycle. GitHub initially pursued general and per-customer fine-tuning, only for rapidly improving foundation models to “Sherlock” that roadmap. Its response is a shared SDK and harness spanning the Copilot desktop app, CLI, cloud agents, security remediation, issue triage, documentation checks, and parallel task execution.
Enterprise AI differentiation is becoming a context-and-permissions problem rather than a model-access problem. The most valuable workflows look backward across PRs, web posts, Obsidian, Teams transcripts, Slack, and email before recommending the next few days of action. Work IQ and Foundry IQ matter because they connect existing corporate stores while retaining enterprise security and compliance controls: “It’s all just context.”
Small, composable skills are proving more durable than monolithic agent workflows. GitHub’s internal teams are “ending the era of these massive beautiful perfect skills,” instead sharing narrow capabilities that can be recombined for each role. The same summarization primitive must produce different judgments for marketing, analysts, customers, or sales—or users immediately ask, “Did AI make this?”
AI disproportionately amplifies leaders who combine technical pattern recognition with expertise accumulated elsewhere. The GitHub guests describe using AI to “crank up 15 agents on Saturday” while their children play lacrosse, then applying years of business knowledge to revenue planning or product creation. That removes low-value slide production but not the chief-of-staff role: roles “don’t all go away, they just change” toward human connection, coordination, and judgment.
Trust, not code generation, may be the binding constraint on agentic development. If one agent writes a PR and another reviews it, “the trust is kind of diffuse”; technical verification still does not answer whether humans, enterprises, and regulators feel safe accepting it. GitHub therefore favors configurable reputation heuristics and costly signals such as sponsorship over universal rules, while acknowledging that stars, account age, and contribution counts are all gameable.
Microsoft’s OpenClaw interest signals a platform bet extending from enterprise context down to operating systems and compute. A useful personal agent needs computer access, but a workplace version also needs OS-level sandboxing so an employee can “use that claw and not get fired.” The opportunity is to supply reusable platform components across Windows, cloud sandboxes, inference, CPU, and silicon—not have every team recreate the same top-level agent product.
Deep dive
Not yet available upstream; scheduled sync will retry.