Sourcegraph vs vLLM

The product line has reorganized around coding agents as the customer. Code Finder runs its own search loop and hands an agent exact files and line ranges; Deep Search now executes scripts in a sandbox against the search APIs so a migration audit returns a CSV checklist rather than thousands of files of tokens; Agentic Batch Changes entered public beta to scope and ship migrations across hundreds of repositories. The argument threaded through all of it is that an agent searching on its own is slower, costlier, and worse than one handed scoped results.

Two consequences of that repositioning are already showing. First, Sourcegraph is pushing into other vendors' agents rather than waiting for its own surface — reranking Claude Code's file picker on symbols instead of path characters is integration work on someone else's tool. Second, scoped retrieval is being sold as governance: if you can prove which files an agent read before it shipped a change, agent context becomes an audit trail, which is the wedge into regulated engineering teams. The security posts extend the same logic from migrations to vulnerability remediation across ten thousand repositories.

Expect Agentic Batch Changes to move from public beta toward general availability, and more retrieval integrations into third-party coding agents. The provenance angle looks like the basis of an enterprise tier aimed at regulated buyers.