I Interviewed an AI Version of GitHub's COO — Then Spoke to the Real One
Mike Taylor (head of tech consulting at Every, co-author of Prompt Engineering for Generative AI) ran a pre-interview experiment before sitting down with GitHub COO Kyle Daigle at Microsoft Build 2026. He built a simulated AI version of Daigle from public writing and talks, asked it the same 12 questions he planned to ask the real Daigle, then compared responses.
Match rate: 2 strong matches, 4 partial matches, 6 material misses. Notably, where the AI lacked evidence it admitted the gap rather than fabricating — and those admitted gaps were the most useful interview prep signal.
The core argument
The experiment is a proof of concept for AI-assisted preparation methodology: build a persona from public signal, surface the known/unknown boundary, then use the gaps as the most targeted interview questions. The article walks through the real interview question-by-question with simulation notes comparing AI-Kyle vs. real-Kyle responses.
Key substance from the Daigle interview (what was accessible pre-paywall):
- Expanding the developer definition — Daigle's art-school-to-developer origin story; GitHub's thesis that Copilot is pulling non-traditional builders (legal, finance) into development workflows
- PR flood management for maintainers — Copilot code review is now agentic; agentic merge and contributor acceptance controls are the near-term answer to agent-generated PR volume
- Agent-generated activity scale — 1B commits in all of 2024; on pace for 14B in 2025 (14× growth); 17 million agent-generated PRs in March 2026 alone
- Business models for always-on agents — paywalled, but secondary sourcing (BigGo) indicates Daigle's framing: "malleable trust heuristics" replacing centralized verification as agent-written code reviewed by agents makes traditional social trust signals obsolete
- Infrastructure strain — three identified bottlenecks: legacy MySQL permissioning at 200M account scale, CPU constraints for CI/CD at machine-speed commit volume, monorepo storage. Daigle projected step-change improvements within three months (by ~Sep 2026)
- Copilot strategy pivot — abandoned fine-tuning investments when base model improvements (GPT-4, Claude 3) made custom tuning obsolete; shifted to universal SDK + multi-model support + agentic runtime
Why this is in the vault
Two distinct value layers:
The interview methodology itself — AI persona simulation as pre-work is a reusable research pattern. The 2/4/6 match distribution is an honest empirical calibration of what public-signal AI personas can and cannot predict. The gaps-as-signal inversion (the misses are the most valuable prep) is a transferable heuristic for any high-stakes conversation.
The GitHub data as infrastructure signal — 14× commit growth driven entirely by agents is not a developer-productivity story; it is an infrastructure and trust architecture story. The "when robots review robots' code" problem is the key unsolved challenge Daigle names. This has direct implications for anyone building agent pipelines that produce artifacts reviewed by downstream agents.
Mapping against Ray Data Co
Strong alignment on the COO-agent thesis. The experiment Taylor ran — build an AI simulacrum of a human executive from public signal, use it to surface known/unknown boundaries before a high-stakes interaction — is a live instance of what RDCO's always-on COO agent should be able to do for Ben's client and prospect preparation. Taylor's 2/6 miss rate is a realistic calibration: AI persona simulation works for public-signal-dense individuals, degrades on novel or private reasoning.
The agent trust architecture problem maps directly to RDCO's multi-agent pipeline design. When the COO agent spawns sub-agents to research, write, and review outputs, and those sub-agents produce artifacts the parent agent then evaluates, we hit exactly the "agent writes + agent reviews = trust diffuses" problem Daigle describes. The independent-worker verification pattern (/verify-vault-write, /verify-dispatch) is RDCO's current answer — but it's worth stress-testing against the "malleable trust heuristics" framing.
Caution on the prep methodology: Taylor's experiment works because Daigle has extensive public signal. For targets with thin public footprints (early-stage founders, private-company operators), the AI persona will have a much higher miss rate and the admitted-gaps signal will dominate. The methodology needs a confidence-floor check before treating gaps as meaningful vs. merely absent.
The 14B commits / 17M agent PRs data is useful context for pitches to any GitHub-dependent client. Agents are no longer a pilot program on GitHub — they are the primary workload. Clients asking "should we try AI coding agents?" are behind the curve; the question is now "how do we manage the agent-generated PR queue?"
⚠️ Sponsorship
None. No third-party sponsors identified. The only promotional block is Every's own subscription paywall upsell (in-house product, not an external advertiser).
Related
- [[06-reference/2026-06-02-every-eight-levels-ai-adoption]] — same author (Mike Taylor); the eight-levels framework maps directly to where the COO-agent stack currently sits (levels 6–8)
- [[06-reference/2026-04-09-every-four-ai-agents]] — Every's own 25-person company running on four AI agents that collectively replace a COO function; the most direct operational parallel to RDCO
- [[06-reference/2026-04-08-every-half-agent-now]] — agent-integrated ops and the trust-via-public-execution insight
- [[06-reference/2026-06-07-alphasignal-async-agents-release-bottleneck]] — human developers as release bottleneck; complements the agent-PR-flood problem Daigle describes