Salesforce finishes migration to Claude Code, removes token limits after 231‑person‑day project

Enterprise coding AI usually gets pitched as a helper. A faster autocomplete. A smarter chatbot in the IDE. But Salesforce is describing something bigger — a shift where the agent is no longer sitting off to the side, and is instead driving chunks of the software delivery process itself. On May 12, Salesforce said it finished a broader internal move to standardize engineering on Anthropic’s Claude Code and removed token limits for its engineers, framing that as the inflection point that changed how work gets done. ### What actually changed? The concrete change was not just “we use AI more now.” Salesforce says it rolled out Claude Code across its engineering organization as the primary AI agent tool, then removed token limits so engineers would stop rationing usage. That matters because token caps don’t just control cost — they change behavior. If developers feel every prompt is expensive or scarce, they use the tool for tiny tasks. Remove that friction, and the tool starts getting used for end-to-end work. (salesforce.com) ### Why are token limits such a big deal? Because limits quietly train people to think small. A capped tool becomes something you consult. An uncapped tool becomes something you delegate to. Salesforce’s argument is basically that the workflow changed once engineers no longer had to budget prompts and context windows like airline miles. The company ties that change to “agentic” use across the software development lifecycle — writing code, reviewing pull requests, generating tests, updating docs, and helping with deployments. (salesforce.com) ### What’s the headline example? Salesforce points to one migration inside its Agentforce Commerce team. The job was to move 33 API endpoints to a new cloud-native architecture. By Salesforce’s estimate, doing that the old way would have taken about 231 person-days, or roughly seven days per API. The team says it finished in 13 days instead — about 18 times faster. That’s the stat doing most of the work here, because it turns vague AI optimism into a specific engineering case study. (salesforce.com) ### Was it just faster, or also better? Salesforce says both. In the same May 12 post, it said work items completed per developer in April 2026 were up 50.8% from April 2025, pull requests merged per developer were up 79%, and its internal “Effective Output” score was up 151.3% year over year. Just as important, Salesforce says quality moved in the right direction too — more shipped, fewer incidents, fewer bugs. That combination is the part companies care about most, because speed without quality just creates cleanup work later. (salesforce.com) ### Why Claude Code specifically? This didn’t come out of nowhere. Salesforce and Anthropic expanded their partnership in October 2025, and both companies said then that Salesforce was deploying Claude Code across its global engineering organization. The broader partnership also made Claude a preferred model for Agentforce in regulated industries and deepened Claude integrations with Salesforce and Slack. So the May 2026 post reads less like a surprise launch and more like the first detailed internal results from a strategy that was already underway. (salesforce.com) ### What does “agentic” mean here? Basically, the model is being treated less like a code assistant and more like a teammate that can carry a task across steps. Not just “write this function,” but “map the schema, generate tests, update docs, review the PR, and help move the change toward deployment.” The catch is that this still depends on governance, measurement, and review. Salesforce itself says the earlier phase was about getting thousands of engineers to adopt AI with the right scaffolding before this deeper shift could happen. (anthropic.com) ### So what’s the real takeaway? The interesting part is not that Salesforce likes Claude Code. Lots of companies like lots of tools. The interesting part is that a giant software company is saying the bottleneck was not model quality alone — it was workflow design, organizational permission, and removing usage friction. If that holds up beyond showcase projects, the new default in enterprise engineering may be fewer prompts, more delegation, and AI agents doing the boring middle of software work by default. (salesforce.com)