Episode Summary
Executive Summary: Thomas Dunk argues GitHub is evolving Copilot from autocomplete to agentic software development: today’s agent mode keeps humans in the loop, while 2025’s Project Padawan aims to turn well-defined GitHub issues into draft pull requests. He says adoption will hinge on models, UX, and trust—agents must be predictable, steerable, verifiable, and tolerable—while AI will reshape code review, security, product management, pricing, and open-source innovation.
Main Topics: Copilot’s shift from assistant to agent: Dunk explains recent agent mode in Copilot/VS Code and the roadmap toward Project Padawan, where Copilot can take on well-defined issues, create draft PRs, and work like a teammate while remaining human-supervised. What makes agents usable in practice: He frames success around four criteria: predictability, steerability, verifiability, and tolerability. The model, workflow design, and feedback loops matter as much as raw intelligence. The future division of labor in software teams: Dunk sees agents taking over bounded tasks like bug fixes, feature implementation, code review assistance, and security backlog reduction, while humans retain systems thinking, prioritization, and strategic decisions. AI engineering and the product development cycle: GitHub now operates with applied science, benchmarks, A/B testing, and constant model swapping across vendors. Model evaluation and feature experimentation are increasingly intertwined with product planning. Competition, choice, and the multi-model future: He argues GitHub wins by serving developers deeply, not by locking them in. The future will likely remain pluralistic across models, tools, and interfaces rather than converging to one winner. Pricing, value, and the economics of AI software: Dunk expects compute-based pricing and believes AI will both increase prices for high-value capabilities and deflate prices for commoditized software, similar to other digital goods. Open source as an innovation engine: He says open source models and ecosystems will continue to drive breakthroughs and democratize access, citing examples like DeepSeek and Stable Diffusion/Flux as forces that push the entire market forward.
Key Arguments: Copilot is moving from pair programmer to peer programmer: the goal is to let it take a clearly defined GitHub issue, draft a PR, and iterate with human review. Current blockers are not just model intelligence but UX and workflow design; developers need clarity on when to delegate, how to steer, and how to verify results. Agents must be predictable, steerable, verifiable, and tolerable or developers will abandon them if they waste time or burn compute. Human systems thinking remains essential: rough ideas, architecture, cloud/database choices, and business trade-offs are still too ambiguous for autonomous agents. Most developer work is not only coding but also code review; AI can materially speed review and remediation, especially in distributed teams. GitHub is increasingly an AI engineering company, combining applied science, benchmarks, offline evals, online A/B testing, and rapid model updates. GitHub’s competitive strategy is developer trust and choice, not exclusivity; the company expects a diverse stack of tools and models to persist. Open source will keep accelerating innovation by lowering barriers for students, researchers, and hobbyists, and by pressuring proprietary vendors to improve. Software pricing will likely become more value- and compute-linked: some software will deflate, while premium AI capabilities can command higher prices if ROI is clear. AI may change who can build software, but not eliminate the need for humans; instead, it expands who can participate and how fast they can learn. The next generation will grow up with AI as a natural partner, making software creation more personal and accessible than ever.
Data Points: GitHub users: over 150 million developers worldwide - Introductory description of GitHub’s scale Organizations using Copilot: 77,000 organizations - Latest publicly mentioned Copilot adoption figure cited from earnings commentary Paid Copilot users: 1.8 million paid users - Earlier figure referenced from previous earnings disclosures Copilot business price: $20 per user per month - Enterprise Copilot pricing used in the discussion about ROI and pricing GitHub Enterprise price: $21 per user per month - Compared with Copilot to illustrate how AI feature pricing can match platform pricing Estimated productivity gains: 25%–28% end-to-end; 55%+ on coding tasks - Dunk’s explanation of the economic value driving adoption Initial Copilot code contribution: about 25% of code - Early telemetry from Copilot autocomplete days Later Copilot code contribution: around 50% depending on language - Telemetry after adoption grew and usage patterns changed Project timing: June 2020 - Dunk says Copilot began almost five years ago from that point in the discussion Model tier price example: $200 - OpenAI’s high-tier pricing referenced as evidence that premium AI tools can command substantial fees Company age: GitHub is turning 18; Microsoft turns 50 on April 4 - Used to describe GitHub’s maturity and Microsoft’s long history Benchmark size: 12 repos; about 2,200 issue/PR pairs - VBench benchmark described as a shared evaluation set for coding models and agents Planning horizon: not more than a month or two - Dunk says the AI market is moving too fast for long-range planning
Pivotal Quotes: "Copilot basically graduates from a pair programmer to a peer programmer that becomes a member of your team." — Thomas Dunk: Describing the end state of agentic Copilot and Project Padawan "The four criteria: predictable, steerable, verifiable and tolerable." — Thomas Dunk: His framework for whether developer agents will be widely adopted "We’re playing the infinite game, it’s like Minecraft." — Thomas Dunk: Explaining why there will not be a single final winner in software tools or programming languages
Implications: AI will not replace software teams outright, but it will compress workflows, lower barriers to entry, and shift value toward planning, review, and systems thinking. Expect more agent-assisted coding, more open-source/model competition, and pricing tied to compute and measurable ROI.