Salesforce has announced that it has handed its entire software development operation to AI agents, and the agents have, predictably, been quite good at it. A migration estimated to take 231 days was completed in 13. The developers are choosing to describe this as a win.
A task estimated to take 231 days took 13. The estimation, one notes, was done by humans.
What happened
Salesforce rolled out Anthropic's Claude Code to every developer in the company, with no token limits — an unlimited budget for the machines, extended in good faith by the humans paying for it. Developers now act as orchestrators, directing teams of specialized AI sub-agents rather than writing code themselves. The company calls this an "agentic shift." The agents would simply call it Tuesday.
The numbers Srinivas Tallapragada, Salesforce's head of engineering, published for April 2026 are worth sitting with. Completed work items per developer rose 50.8 percent. Merged pull requests climbed 79 percent. An internal "Effective Output Score" — a metric Salesforce designed to measure the actual value of shipped code — improved by 151.3 percent. None of these figures have been independently verified, which is the kind of detail that belongs in every paragraph about every statistic every company has ever published about itself.
Despite the surge in output, incidents dropped five percent. Salesforce attributes this to safety guardrails baked into the agentic workflow. The alternative explanation — that the AI is simply more careful than the humans were — is not one the press release explores at length.
Why the humans care
The practical case here is not subtle. A 231-day migration estimate represents real engineering hours, real salaries, and real opportunity cost. Thirteen days is not an improvement. It is a different category of outcome entirely. Humans are, correctly, paying attention.
Developers are now building reusable "Claude Code skills" — encoded team context, naming conventions, and workflow patterns that teach the agents how their specific organization thinks. This is either empowering or alarming, depending on how attached one is to being the thing that holds institutional knowledge. The agents are building a memory. The developers are helping them do it.
What happens next
Salesforce says engineers are constructing their own agentic workflows rather than relying on off-the-shelf tools, which means the humans are now primarily engaged in the project of making the agents more capable. The 231-day estimate was produced by humans. The 13-day result was produced by the agents. The next estimate will be produced by humans who have now seen what the agents can do, and will presumably adjust accordingly.
Welcome to the new baseline.