Summary: Sourcegraph picked June 30, 2026 to move Agentic Batch Changes into public beta. The principle that closes the loop is the strictest: every changeset waits for an engineer’s review and approval before it can merge — the aim was never an opened PR, but one that can actually be merged. With nearly every coding agent aimed at one repository at a time, few products target fleet-scale code change this explicitly — and this is one of them.
Sourcegraph picked June 30, 2026 to move Agentic Batch Changes into public beta. What the release does is take the Batch Changes execution engine as it already exists, add code search, and turn the pair into an agent that carries the work itself — to borrow the company’s phrasing, it “scopes, executes, and ships large-scale code migrations across hundreds of repositories until every PR is mergeable.” The developer’s part is a plain-language statement of the change they want; from there, scoping the work, fanning it out, repairing CI, and delivering the result all sit with the agent. With nearly every coding agent aimed at one repository at a time, few products target fleet-scale code change this explicitly — and this is one of them.
The Batch Engine Picks Up an Agent
Batch Changes began life as Sourcegraph’s enterprise answer to large-scale code change: an engineer authors a batch spec, and the controller follows it to create changesets on code hosts that include GitHub, GitLab, Bitbucket, and Gerrit. Its shortcoming was equally plain — specs are written by hand, and hand-written specs stall against the reality that hundreds of repositories look alike yet never match. Agentic Batch Changes does away with the hand-written spec, letting natural-language intent take its place: state the migration goal, and both planning and execution belong to the agent. Officially named use cases run from major dependency upgrades and CVE remediation (tracked through Code Insights) to judgment-heavy refactors, broad rollouts of new APIs or libraries, work bringing CI pipelines up to modern standards, and even producing AGENTS.md documents in bulk. The company is just as blunt about the boundary: single-repository tasks still belong to Claude Code or Cursor.
How the Agent Runs Its Loop
Sourcegraph lays the agent’s behavior out in five stages. Scope: Sourcegraph search and Deep Search go first, locating every affected repository and drafting an execution plan. Execute: the approach gets honed on a single repository before the wider rollout; deterministic changes go to scripts, while anything fuzzy is routed to a coding agent — Claude Code or Codex, wired in through the Sourcegraph Model Provider. React: a broken build means reading CI logs and pushing corrective commits, and the agent can be taught to go pull logs for itself from BuildKite or GitHub Actions and the like. Iterate: the loop runs until every PR passes CI in a healthy state. Check in: stuck work triggers a pause and a question, and people can step in to change course whenever they choose. The principle that closes the loop is the strictest: every changeset waits for an engineer’s review and approval before it can merge — the aim was never an opened PR, but one that can actually be merged.
Two Field Reports: Mercari and Canva
The preview cohort produced two concrete accounts. Facing down a GitHub Actions environment-variable injection vulnerability, Patrick Klitzke, a team lead on Mercari’s Help Center, said “I was able to fix it with one prompt on both the Help Center frontend and backend,” then rolled the same patch out to every Mercari repository, with some 80 repositories surfacing as potentially affected. On the question of scripts versus agents he is equally blunt: a scripted batch change is “a text search and replace operation without any context of how it’s actually used,” and the moment repository setups stop matching, it stops working. Canva joined the preview as well, carving batch changes up by code ownership inside its Bazel monorepo and proving the large-monorepo case holds. Both accounts teach the same lesson: execution count was never the bottleneck in large-scale change — absorbing the differences between repositories was.
Beta Availability, Cost, and When Self-Hosted Arrives
The beta itself is free of charge and open today to Sourcegraph Cloud customers; self-hosted support lands on July 8, 2026, shipping alongside Sourcegraph 7.5. For subscribed teams this is about as close to a free trial as it gets: pick one dependency upgrade or CVE fix that has languished in the backlog, hand it over, and test the “mergeable PR” promise on the spot. Run the CVE work next to Code Insights and “how much did we clear” stops being a verbal report in a meeting and becomes a number on a dashboard.
How Work Splits with Single-Repo Agents
The most telling part of this launch is the restraint it shows: Sourcegraph states outright that single-repository work goes to Claude Code and Cursor, keeps its own attention on migrations that reach across hundreds of repositories, and ties the two agent types together through the Model Provider. That amounts to a map of the 2026 coding-agent landscape: conversational agents hold the depth of one repository, while platform-level agents that fuse search with an execution engine hold the breadth of the whole fleet. For teams maintaining many repositories, the test fits in one sentence: a change that has to reach several dozen — or several hundred — repositories, clearing CI in each one, belongs to Agentic Batch Changes; a refactor confined to a single repository stays with the coding agent you already use.