Sourcegraph's Merge Pricing Leaves Review on Your Tab • The Weighted Average

Sourcegraph made Agentic Batch Changes generally available on September 14, moving its coordination agent to pricing based on merged changesets. That comes 76 days after its public-beta announcement—a commercial transition that fleet owners should review before treating yesterday’s free experiment as tomorrow’s migration budget.

The bill moves to the merge button

The product addresses a specific engineering burden: carrying one intended change across repositories that resemble each other without being identical. A coding agent can alter a service; a fleet migration must also identify every affected service, adapt the change, handle failed checks, and show which repositories remain unfinished. Sourcegraph sells that coordination layer. The near-term buyer is a platform team with repeated cross-repository work, not an individual looking for a different autocomplete tool.

The calendar calculation combines two first-party records. Sourcegraph’s June 30 public-beta announcement said usage was free during beta. Its September 14 release establishes general availability and outcome-based pricing. The elapsed interval is 76 days: 31 days in July, 31 in August, and 14 in September. That measures the interval between announcements, not continuous uptime, customer exposure, or an independently audited trial period.

This distinction matters because the announcement reports nearly a thousand merged beta changesets, but provides no denominator of attempted changes or reviewer hours. A merged total establishes that customers used the system. It cannot establish a success rate or labor saving. Nor should it be divided into traditional Batch Changes’ roughly half-million annual merges to claim market share: those figures cover different products and time windows.

The current product page says customers pay per merged changeset rather than per token, seat, or attempt. The launch release is more specific about scope: the outcome-priced component is the coordination agent. That is a useful change in incentive, but not evidence that every part of a migration becomes free when a pull request is rejected. Obtain a written account of coding-agent charges, executor resources, CI, and existing platform entitlements before comparing offers.

Sourcegraph does not publish a numerical per-merge tariff in these retrieved launch materials. That missing price is a procurement question, not permission to invent one. Ask what counts as a billable merge, how a reverted change is treated, and whether splitting a migration into separate changesets changes the charge. These are questions the agreement should answer; the announcement does not establish the answers.

The customer example makes the unit concrete. Sourcegraph quotes Canva engineer William L. describing 50+ pull requests raised and merged for a library migration, with a shared interface replacing large pull requests or spreadsheets. That supports a coordination use case. It does not reveal Canva’s invoice, compare acceptance quality with another tool, or quantify how much review disappeared. The strongest defensible benefit is visibility across the rollout.

A ready pull request is still somebody’s responsibility

The operating documentation describes a staged workflow: research the scope, approve a plan, test one repository, inspect the diff, publish a draft where supported, and expand. The agent can react to CI failures when configured to retrieve the necessary logs. It cannot merge changesets from the conversation; engineers merge through the code host or the supported Sourcegraph interface. The control boundary is explicit, and teams should preserve it.

There is retained setup work. Users need publishing credentials for each code host. CI must report checks to that host, and fetching logs requires appropriate access. A pilot should test these connections before launching the broad change. Otherwise the team may confuse an unavailable log or missing permission with a coding failure and spend review time diagnosing infrastructure that was never ready.

The routing mechanism deserves equal attention. Sourcegraph says the coordinator uses deterministic scripts for mechanical changes and delegates judgment-heavy work to coding agents such as Claude Code or Codex. That is more sensible than asking a model to rediscover an identical edit in every repository. But the operator still needs to verify which differences are mechanical and which are semantic. An elegant script can distribute a mistaken premise with exceptional consistency.

The archive’s Fugu analysis separates orchestration prices from accepted-task costs. Sourcegraph changes the billable boundary, not that accounting principle. Keep a migration ledger containing the target repository set, completed merges, exclusions, rejected changes, reviewer time, and remaining obsolete usage. A smaller invoice is not success if the supposedly completed migration leaves important callers untouched.

Today’s Temporal lead examines the cost of keeping long-running work recoverable. Here the analogous question is whether the rollout can stop, resume, and explain its state without an engineer rebuilding the spreadsheet. Test interruption and revision as well as the happy path. Ask the system to account for what remains, rather than rewarding the number of pull requests it managed to open.

The strongest counterargument is that a mature scripted migration already does enough. Sourcegraph’s own beta announcement directs single-repository work toward ordinary coding tools. Teams with uniform repositories and reliable existing automation should not add a coordinator merely because it now has an attractive billing label. The product earns its place where variation and tracking consume material effort that the existing approach cannot cheaply remove.

Evidence that would change the decision is a representative rollout with lower total cost per accepted migration, complete scope coverage, and no deterioration in review quality or rollback burden. Choose a reversible library or configuration change, preserve the incumbent approach as a baseline, and include unsuccessful attempts. Adopt for demonstrated coordination savings; negotiate before scaling. Pay-per-merge is a cleaner meter than pay-per-attempt, but review and responsibility remain on the buyer’s side of the counter.