Continuous AI (with Don Syme)

Notes and transcript of an AI Native Dev podcast episode (hosts Guy Pigiani and Simon Maple) featuring Don Syme, principal researcher at GitHub. Syme introduces Continuous AI as a distinct category alongside CI and CD, and details GitHub Agentic Workflows, GitHub's implementation of it, now in public preview.

The core idea

CI establishes that software works; CD establishes that it is correctly deployed — both deterministic. Continuous AI covers the subjective activities — bug triage, documentation, performance improvement, duplicate detection, issue labeling — that need to run on a permanent, operational basis inside a bounded context, but have no binary pass/fail outcome. Syme is emphatic that this is not AI-inside-CI: "It's CI CD, they stay as they are."

The shift that matters, Syme argues, is from individual productivity (the chat modality) to situated, collaborative automation — AI folded into the change/event-driven rhythm that already governs software teams, with the repository as the natural bounded context.

Guardrails, review, and cost

Syme's running metaphors: "the stronger the train tracks, the faster you can run" — guardrails are an enabler of speed, not a limit — and "you've got to be able to sleep at night." Information-flow integrity (never expose private repos to public-repo agents, default to ignoring untrusted contributors) is existential.

Human review is the bottleneck, so the factory creator's job is to equip the reviewer: evidence (before/after CI timings), risk assessment, trade-offs, and quality gates that throw away trash before it reaches a human. "If it produces rubbish, you throw it away."

Cost control, Syme claims, is "the thing that's missing from the harness discussion." GitHub Agentic Workflows ships per-run cost visibility, budgeting, and "examination testing" — running the same work with multiple models simultaneously to compare performance.

The workflow debate

Pelle de Hallio's "Agent Zoo" explores dozens of workflow patterns. But for maintaining real repos, Syme prefers one broad-spectrum workflow (his own RepoAssist: labels issues, does first-response research, optionally implements fixes, never auto-merges) for conceptual simplicity, cost control, and attention management across many repos.

Open questions

The transcript leaves human-review scaling (auto-merge pressure), evaluating subjective automation quality, cross-repo/org context, and prompt-injection security largely unaddressed — Syme acknowledges them as future work.