Sourcegraph GA: keep-working batch agent until CI is green

Sourcegraph’s Sep 14 GA for Agentic Batch Changes is a tortoise receipt at fleet scale. The problem they name is familiar: a change that must land everywhere (EOL bump, CVE, deprecated API, shared-library rollout) turns into hundreds of PRs nobody staffs — while coding agents keep adding code faster than any human can vouch for a codebase-wide migration. Classic Batch Changes already opens and tracks PRs across hosts; the agentic layer plans the work, adapts per repo, and keeps working until the pull requests are ready for an engineer to review and merge.

Loop shape (from the announcement and product page): engineer states the change in plain language → coordination agent scopes via Sourcegraph’s indexed code graph → validates the approach on one repo before blasting diffs → pushes in batches → routes mechanical work to scripts and judgment work to Claude Code/Codex (with per-repo instructions + Sourcegraph MCP) → on CI failure, adjusts and retries → humans approve every changeset before merge. Outcome-based pricing charges per merged changeset, not per token or abandoned PR — an economic nudge toward verified throughput over bursty attempt volume.

Beta receipts they cite: nearly a thousand changesets merged on security remediation through library migrations. Mercari’s Patrick Klitzke: one prompt fixed a GitHub Actions env-injection pattern on Help Center FE/BE, then extended across ~80 affected repos — “similar, but not identical setups” that a blind search-replace would mishandle. Canva’s William L.: 50+ PRs for a library migration, tracked to merge from one UI instead of “huge PRs or spreadsheets.” Batch Changes baseline they quote: half a million changesets/year already; GA adds hosts GitHub, GitLab, Bitbucket Server/Cloud, Azure DevOps, Gerrit for Sourcegraph Cloud.

Why it matters for code review

Steady harnessed momentum compounds past peak-session heroics: the scarce resource is still human review/merge judgment, but the factory’s job is to present CI-green, scoped, trackable changesets overnight and across the fleet — not a hero Claude session in one repo. FOMO when agents aren’t working on your behalf is exactly the deferred migration backlog Adler names; the counter-risk is PR flood without verification. Sourcegraph’s bet is explicit: agent loop owns CI repair and cross-repo variation; humans keep the merge button. If your “always-on” story stops at generation without a keep-working-until-green gate, you bought the hare’s FOMO and still pay the morning review tax.