Source: raw/Pants-based_Benchmarks_-_GitHub_s_Agent_Revolution.md — Creator: Intelligent Machines (TWiT) episode 892, hosts Leo Laporte, Paris Martineau and Jeff Jarvis, guest Martin Woodward, GitHub’s vice president of developer relations · URL: https://www.youtube.com/watch?v=zfw-badCV9s · Platform: YouTube (auto-captions, no speaker labels; recorded 2026-10-14, fetched 2026-10-21). The interview runs from the start of the episode to roughly line 1300 of the transcript; the rest is news.

Woodward gives GitHub’s view of what agents have done to the platform: pull-request volume roughly tripled in a year, open-source maintainers needed new controls against well-meant low-quality contributions, and he argues that teams with good DevOps habits get the most out of agents. He also describes his own working setup, including a writing-voice skill built from 20 years of his blog posts. The platform figures are his own and unaudited, and his claim that Copilot is the only harness with every lab’s models is a vendor claim.


Key Takeaways

  • PR volume, per GitHub. “In August … across GitHub … 130 million pull requests merge per month and if we look a year earlier it was around about 40 million”, and “like 50% of those … are kind of being agent assisted.” The hosts add that some users report “2,000 PRs every day.”
  • November 2025 was the turning point. With the new Claude Sonnet and OpenAI models that month, “the amount of time that agents can code without human interaction just increased exponentially.” Leo notes GitHub’s downtime rose under the load; Woodward says it has “got a lot better”.
  • No job-loss signal in GitHub’s data yet. “We’re just not seeing it in the data … the number of developers going up and up”, and first-time coders are “rising massively”.
  • Maintainers now have off switches. Low-quality contributions are “not really spam because often times it’s well-intentioned. It’s just not good.” GitHub added the ability to switch off pull requests, temporarily or for people who have not contributed before, and to cap how many unaccepted changes one person can submit.
  • Fight the slop with automation. Run tests, code checks and security checks in GitHub Actions before accepting a change, then have an AI reviewer flag things such as new dependencies added “sneakily” by code that “actually … steal[s] crypto wallet from you”.
  • DevOps habits transfer to agents. “Putting the kind of rules in place to do DevOps well for humans also means that you can do AI better … People who are good at DevOps tend to be good at using AI.” And: “if you put slop in, you get slop out.”
  • Constraints produce good output. The GitHub Universe badge ships an agents.md with its pinout and MicroPython version: “give the agent the appropriate skills and constraints and … wonderful things happen.”
  • Use AI to implement, not to think. “I actually use AI to help me implement. I try not to use AI to help me think … the process of writing is the process of thinking.”

How Woodward works

  • Three surfaces plus skills. A CLI, an app and the IDE, with a set of agent skills. The IDE is for “the fine areas” and the agent for “the broad strokes”, like an oil painting.
  • Background sessions. He starts parallel coding sessions before bed or during a dull call; a hacked Furby made noises when Copilot finished.
  • A voice skill from his own archive. He had his agent review “20 years of blog posts” and build a skill to “sound like me”: British English, no em dashes, his habitual words. The agent derived the rules from his writing and applies them when helping him draft.
  • Several model families in one harness. Copilot offers Anthropic models (Sonnet, Opus, Fable), OpenAI models (GPT-5.6, Sol, Astra, Terra) and open-weight models (DeepSeek, Kimi). The guest describes having one model plan, a different family code and a third review “so that they … have overlap with their gaps”. Saying “rubber duck this for me” in Copilot makes it consult another model. “We’re the only ones that’s got all the models … in one harness” is his claim, not a verified comparison.
  • Check understanding first. A tip from the conversation: ask the model “tell me what you understand from what I just told you”, because it often says “something nonsensical” that you can correct before it starts.
  • Personal, ephemeral software. New application types include “an application which just solves a problem for one person”, which nobody would have coded before because “you only [code] when you do something twice”.

Try It

  1. Build a writing-voice skill from your own archive. Point an agent at your past posts or emails and ask it to list the rules it observes (spelling variant, banned punctuation, recurring words), then save them as a skill. For a brand, use approved published copy as the archive.
  2. Ask for a read-back before a long task: “tell me what you understand from what I just told you.”
  3. Split plan, code and review across model families, as described in the episode; see the same habit in DHH on Lex Fridman.
  4. If you maintain a public repo, set contribution limits (pull requests off for first-time contributors, a cap on open unaccepted changes) and run checks and an AI dependency review in Actions before human review.
  5. Write an agents.md with the hard facts of your environment (hardware, versions, conventions) rather than general advice.

Open Questions

  • GitHub’s 130 million and 40 million figures and the 50% agent-assisted share are a spokesperson’s numbers from memory on air; no published dashboard was cited.
  • Auto-captions carry no speaker labels, so attributing the plan/code/review routine and the read-back tip to Woodward or Leo is uncertain.