Every Repo Is a Software Factory Now | Don Syme, GitHub

GitHub Principal Researcher Don Syme describes “continuous AI”: recurring agent work such as documentation, issue triage, and performance research, shared by a team rather than confined to individual chats. In this AI Native Dev interview, he argues that these judgment-based tasks should run beside deterministic CI/CD, with quality checks that keep weak proposals from consuming human review time. He draws on his work on GitHub Agentic Workflows, then in public preview.

Keep Subjective Automation Beside Deterministic Delivery

• Continuous AI moves agent work from one-off chat to shared operations: Syme defines it as recurring, context-aware automation for subjective work such as documentation, issue triage, performance research, and bug fixing. A useful result must become an operational process that keeps responding to repository change, not remain a successful demonstration 03:25–08:32(opens in a new tab).
• Do not turn CI/CD itself into a subjective agent loop: CI/CD should continue establishing deterministic invariants such as a passing build and a correct deployment. Continuous AI is a separate lane that teams can adopt and scale independently while using CI/CD as a quality gate 09:26–12:45(opens in a new tab).
• Bounded context is the first reliability control: Syme recommends constraining each run to explicit inputs, a repository context, and narrow outcomes such as one issue or pull request, with a firewall around the run. His operational test is whether the automation can safely run while maintainers are asleep 09:26–10:16(opens in a new tab), 18:42–21:21(opens in a new tab).

Treat the Repository as a Boundary, Not a Universal Answer

• GitHub Agentic Workflows applies the model through GitHub Actions: Syme describes a public-preview system that runs a choice of coding-agent CLIs and models under Actions, with explicit controls around credentials, inputs, and effects. For example, workflows ignore issues from untrusted contributors unless maintainers opt in to processing them 18:42–22:12(opens in a new tab).
• The repository is also an organizational permission boundary: CI/CD succeeds partly because organizations let repository maintainers consume bounded compute, network, and storage without requesting infrastructure for every automation. Syme expects continuous AI to use a similar delegation model, supplemented by policy, accounting, and cost visibility 25:30–31:34(opens in a new tab).
• Repo-local execution works best when the relevant data is already there: Shared skills and MCP services can supply reusable behavior or external context. A side repository can own authorization for automation targeting another codebase, while specialized systems may remain outside GitHub and use a workflow only to report findings back into the development ledger 37:32–44:36(opens in a new tab).
• Monorepos and cross-repository systems weaken the boundary: Consolidating code helps current agents find context, but a large monorepo may be too broad for clean access, security, and ownership decisions. Syme says GitHub still needs better multi-repository and organization-level context; the interview does not present a finished design 32:26–35:49(opens in a new tab).

Make Cost, Evaluation, and Human Attention Explicit

• More workflows create more maintenance and attention costs: Syme contrasts an experimental “agent zoo” with his preference for one broader repository assistant that can label, research, and sometimes fix issues. Scheduled batching, adjustable cadence, model delegation, and per-run cost history make the operational load easier to see and control 45:32–50:32(opens in a new tab).
• Evaluate models concurrently against the same work: Because a later model can discover another model’s issue or pull-request result in the shared ledger, GitHub’s described “exam testing” starts multiple models at the same time and accepts only one result. The episode explains the contamination risk but reports no resulting scores or selection data 50:32–51:23(opens in a new tab).
• Human approval remains the product’s code-change boundary: Syme says agent-created pull requests are not automatically merged, and GitHub limits workflows that try to modify their own automation. A workflow may make constrained follow-up pushes to its own pull-request branch, but a person still approves changes to the codebase 53:56–58:19(opens in a new tab).
• Quality gates should protect reviewer attention: The target is not more pull requests but reviewable proposals that explain motivation, risk, tradeoffs, and evidence. Syme’s CI-performance example compares before-and-after timings so a reviewer can verify the claimed improvement; low-quality or unsupported output should be discarded before it reaches a person 59:55–1:02:29(opens in a new tab).

Watch the full interview on YouTube(opens in a new tab).