GitHub lets repos set confidence thresholds for agents

GitHub added confidence thresholds to its Issues agent automation on July 23, per the company's changelog.

Repo admins now pick a number instead of trusting a default. Too loose, and unreviewed agent changes pile up on a busy repo; too tight, and nobody turns the feature on at all.

The company calls the control an automation level: admins set it once, and it decides which agent changes ship immediately versus which get held. GitHub's framing: "a small team moving fast might let the agent run on its own, while a busy public repository can hold changes for review."

Three workflows ship in this preview: triage, metadata enrichment, and spam detection. Spam detection shows the pattern best. It flags a change with a reason attached and holds anything uncertain, so a maintainer sees why before acting instead of trusting a black box.

My bet: GitHub sets the default threshold loose to keep adoption numbers up. The first public misstep lands in spam detection, not triage, when an uncertain flag ships without review.

Whether that default is the same across all three workflows, the changelog doesn't say. The preview posted July 23, giving admins only two days so far to find the answer the hard way.