Scheduled agents need run cards

GitHub's June 2 changelog says Copilot cloud agent automations can run hourly, daily, weekly, or in response to new issues and pull request activity. GitHub gives examples like triaging incoming issues, checking for failing tests each night, attempting a fix, and drafting release notes on a schedule.

The important detail is scope. Each automation is tied to a single repository where the agent can read and write code, open pull requests, and update issues. The creator chooses the prompt, trigger, available tools, and model. Token usage is billed to the user who created the automation.

That is already enough to change team behavior. A repo can now contain small agents that wake up before standup. If the run is useful, nobody cares that it was boring. If it is wrong, everyone suddenly cares who configured it.

GitHub's June 17 Copilot post is less flashy, but it may matter more for daily use. Copilot is expanding Auto model selection so the system can choose a model based on task intent and current model health instead of asking the developer to pick every time.

GitHub says its HyDRA router looks at reasoning depth, code complexity, debugging difficulty, and tool orchestration needs. It also says no single model performed best across every task in its evaluations. Stronger models helped most when the work demanded deeper reasoning; cheaper or faster models often reached the same outcome on simpler work.

For teams, the nice version is obvious: fewer model-picker debates. The annoying version is also obvious: if Auto picks the route, the run card needs to say what happened. People do not need every routing score. They do need enough to explain a slow run, a surprising bill, or a weak result.

OpenAI's April Codex update pushes the same direction from another angle. Codex can now operate a computer on macOS, use more everyday tools, remember preferences in preview, and schedule future work so it can wake up later and continue a task that lasts days or weeks.

Anthropic's research on roughly 400,000 Claude Code sessions gives the human side of that shift. The study found that people still make most planning decisions, while Claude handles most execution decisions. Anthropic's line is plain: people decide what to build, and the agent decides how to build it.

Claude Code Artifacts, reported by SD Times on June 19, turns a session into a web page teammates can open: a PR walkthrough, incident timeline, dashboard, or release checklist that updates in place and keeps version history. That helps with the inspection problem. It still does not replace ownership. A live page can show what the agent found; the run card says who is accountable for the next move.

I would start smaller than the vendor demos. One scheduled automation. One repo. One boring job with an obvious pass or fail condition. Issue labels, dead-link checks, nightly test triage, or release-note drafts are better first candidates than a broad refactor bot with a heroic prompt.

Before the second automation, write the run card. It should name the trigger, repo scope, allowed tools, human owner, billing owner, model setting or routing mode, output location, evidence links, refusal conditions, rollback path, and expiry date. If nobody owns expiry, the agent becomes old infrastructure with a newer interface.

The run card does not need to be heavy. It can be a markdown file, a pinned PR comment, or a small page in the agent UI. The point is that a teammate can join after the run and understand what happened without interrogating the person who happened to set it up.

Watch how vendors talk about scheduled work over the next few months. The weak pitch will be more autonomy. The stronger pitch will be legible unattended work: clear triggers, contained scope, boring receipts, and output that fits into the team's existing review path.

That is where adoption usually breaks. Not because the agent cannot do anything useful, but because the useful thing arrives in a shape the team cannot trust yet.