On June 30, 2026, Sourcegraph announced that Agentic Batch Changes is in public beta. The new capability upgrades the existing Batch Changes execution engine and code search into an agent that does the work itself: in the company’s words, it “scopes, executes, and ships large-scale code migrations across hundreds of repositories until every PR is mergeable.” A developer describes the change in plain language; scoping, fan-out, CI repair, and final delivery are all handled by the agent. At a moment when most coding agents focus on a single repository, this is one of the few products explicitly aimed at fleet-scale code change.
The Batch Changes Engine Grows an Agent
Batch Changes has long been Sourcegraph’s enterprise feature for large-scale code change: an engineer writes a batch spec, and the controller uses it to create changesets across code hosts like GitHub, GitLab, Bitbucket, and Gerrit. The limitation was just as clear — the spec has to be hand-written, and reality rarely cooperates when hundreds of repositories look similar but never identical. Agentic Batch Changes swaps the hand-written spec for natural-language intent: you describe the migration goal, and the agent plans and executes. The official use-case list includes major dependency upgrades, CVE remediation (tracked with Code Insights), judgment-heavy refactors, rolling out new APIs or libraries, CI pipeline modernization, and even generating AGENTS.md files at scale. The company is equally frank about what it is not for: single-repository tasks remain the territory of Claude Code or Cursor.
The Agent’s Working Loop
Sourcegraph breaks the agent’s behavior into five stages. Scope: it first uses Sourcegraph search and Deep Search to find every affected repository and build an execution plan. Execute: it refines its approach on one repository first, then fans out; deterministic changes get scripted, while ambiguous ones are routed to a coding agent such as Claude Code or Codex via the Sourcegraph Model Provider. React: when builds fail it reads CI logs and pushes follow-up commits, and it can be taught to fetch logs from BuildKite, GitHub Actions, and similar systems. Iterate: it keeps working until every PR is CI-healthy. Check in: it pauses to ask when stuck, and humans can interject to redirect at any time. The closing principle matters most: every changeset must be reviewed and approved by an engineer before merge — the goal is not an opened PR, but a mergeable one.
Field Notes: Mercari and Canva
Preview customers supplied two concrete stories. Patrick Klitzke, a team lead on Mercari’s Help Center, used it against a GitHub Actions environment-variable injection vulnerability: “I was able to fix it with one prompt on both the Help Center frontend and backend,” then extended the same fix across all Mercari repositories, surfacing roughly 80 potentially affected targets. His verdict on scripts versus agents is just as direct — a scripted batch change amounts to “a text search and replace operation without any context of how it’s actually used,” and it breaks the moment repository setups differ. Canva also joined the preview, splitting batch changes by code ownership across a Bazel monorepo and validating the large-monorepo scenario. Both stories point to the same lesson: the bottleneck in large-scale change was never execution count, but handling what differs from repo to repo.
Beta Scope, Pricing, and Self-Hosted Timing
The beta is free, live now for Sourcegraph Cloud customers, with self-hosted support arriving July 8, 2026 alongside Sourcegraph 7.5. For existing subscribers it is close to a zero-cost trial: hand it one long-postponed dependency upgrade or CVE fix and judge the “mergeable PR” claim directly. Pair CVE remediation with Code Insights, and “how much did we clear” becomes a number on a dashboard instead of a status update in a meeting.
Division of Labor with Single-Repo Agents
The most interesting thing about this launch is what it refuses to claim: Sourcegraph openly cedes single-repository work to Claude Code and Cursor, focuses on migrations spanning hundreds of repositories, and connects the two worlds through the Model Provider. That sketches the division of labor in 2026’s coding-agent landscape — conversational agents own depth on one repo, platform-level agents that combine search and an execution engine own breadth across the fleet. The test for teams maintaining many repositories is simple: if a change must land across dozens or hundreds of repositories and pass CI in every one, this is Agentic Batch Changes’ home turf; if it is a refactor inside one repo, stay with your existing coding agent.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.