Salesforce and Claude Code: A Revolution in AI Development

Salesforce Adopts Claude Code to Transform Software Development

Salesforce has made a significant shift in its software development approach by integrating AI agents, notably Claude Code from Anthropic, across all its processes. This transition marks a pivotal change for the company, which now utilizes AI agents without token limits for software creation.

The results of this integration are impressive: Salesforce developers have managed to produce 79% more pull requests while reducing errors. An API migration, initially expected to take 231 days, was completed in just 13 days.

Developers Become Orchestrators

In this new model, developers no longer write code line by line. They now orchestrate teams of specialized AI agents to handle complex tasks. This approach, referred to as "agentic shift," is supported by Salesforce, which has deployed Claude Code company-wide, providing each developer with unlimited tokens.

By April 2026, Salesforce reported a 50.8% increase in work items completed per developer compared to the previous year. The effective output score, measuring the value of shipped code, surged by 151.3%, although these figures have not been independently verified.

Increased Production and Reduced Incidents

Srinivas Tallapragada, head of engineering at Salesforce, emphasized that despite the increase in pull requests, incidents have decreased by 5%. Thanks to the Engineering 360 platform, security safeguards and quality standards are integrated into the agentic workflow.

Salesforce has also developed tools like the "AI Expert Suite" and "Salesforce Foundation Plugins" to standardize and share best practices among developers. Sub-agents and teams of agents enable the management of parallel workflows, simplifying the breakdown of complex tasks.

A Concrete Example of Rapid Migration

A striking example of this efficiency is the migration of 33 API endpoints to a new cloud-native architecture. This project, which would typically take 231 person-days, was completed in 13 days using a rule-based framework and reference implementations.

Integrated feedback cycles improved accuracy, while autonomous LLM loops operated without human intervention. The result was five pull requests, including one major request covering 21 endpoints with full test coverage.

Challenges and Criticisms

Despite these successes, Tallapragada acknowledges several unresolved challenges, including context management during long agentic sessions and the increased security needed for agents operating on systems. The question of how junior engineers evolve in this new framework also remains open.

Critical voices, such as George Hotz, argue that the use of AI agents could be a costly mistake for the industry. Hotz and others point out that current AI models cannot effectively replace human programming and that the generated code may be of inferior quality.

Despite these criticisms, Salesforce continues to explore this path, convinced of the potential benefits of AI in software development.

More Production, Fewer Incidents

The obvious question of whether quality suffers at this pace is addressed by Tallapragada, who refers to the company's monitoring platform, Engineering 360. Despite the increase in pull requests, incidents have decreased by 5%. Security safeguards and quality standards are integrated into the agentic workflow, he says.

"When agentic tools are applied correctly, quality does not suffer from speed. It benefits from it," writes Tallapragada. Salesforce does not support this claim with external audits or independent measures.

Engineers are now building their own agentic workflows rather than simply using standard tools, according to Tallapragada. The so-called Claude Code skills, reusable capabilities that codify team context, naming conventions, and workflow patterns, have become a new type of engineering artifact. Salesforce has also created an organized library called "AI Expert Suite" and "Salesforce Foundation Plugins," which serves as a shared foundation for all developers.

Sub-agents and teams of agents, specialized AI agents managing parallel workflows within a larger task, are changing how complex work is broken down. Developers no longer navigate between five systems. They describe the desired outcome, and coordinated agents manage the individual steps.

API Migration in 13 Days Instead of 231

As a concrete example, Tallapragada cites the migration of 33 API endpoints to a new cloud-native architecture. The traditional approach would have taken about 231 person-days, the company estimates. By using a rule-based framework built on Claude with Markdown files and reference implementations, the migration was completed in 13 days—18 times faster.

Each feedback cycle on pull requests was integrated into the rule set, allowing accuracy to continue improving. Autonomous LLM loops for building, correcting, and validating operated without manual intervention. Migrations were parallelized across isolated environments. The result: five pull requests, with the largest single request delivering 21 endpoints with full test coverage.

"The most important skill today is knowing how to structure problems for an agentic system, when to delegate versus staying involved, and how to build reusable models that your team can rely on," writes Tallapragada.

Security, Junior Talent, and Team Structure Remain Unresolved Issues

Tallapragada is candid about a series of unresolved issues, calling them "really hard." Context management during long agentic sessions is a skill engineers still need to learn. The quality of CLAUDE.md files—persistent context configurations that align Claude with a codebase—varies significantly between teams and has a major impact on output quality. Security also requires new thinking. When agents act on systems rather than simply making suggestions, the scope of a misconfigured tool becomes much broader.

There is also the question of the talent pipeline. "When agents manage more of the execution layer, how do junior engineers evolve into senior engineers if AI absorbs much of the entry-level work? What is the role of a designer or product manager in this new world?" writes Tallapragada. Salesforce is experimenting with units of one or three people instead of traditional Scrum teams. They do not yet have clear answers.

Productivity Boost or Technical Debt on Autopilot?

A very different opinion was expressed a few days ago by well-known programmer and hacker George Hotz. Using AI agents in software development will be one of the industry's costliest mistakes, he argues.

LLMs are "sophisticated statistical models" that "mimic the distribution of programming" but can never truly program, says Hotz. Large organizations are particularly at risk because less experienced developers may not spot defective outputs.

Even Andrej Karpathy, who now counts himself among the proponents of agentic coding, has reported quality issues. The code generated by agents "is not necessarily super amazing code all the time," he said, describing it as "bloated, with a lot of copy-pasting, clumsy abstractions that are fragile, and it works, but it's just really ugly." Unlike Hotz, however, Karpathy remains convinced by the new approach and has recently joined Anthropic.

A broader debate about the rising costs of AI versus its benefits is also intensifying, alongside questions about what models actually deliver in daily work.