A Salesforce product team scoped a migration of 33 API endpoints at roughly 231 person-days of work. Seven days per endpoint: schema mapping, testing, documentation, all the low-leverage toil that keeps engineers busy without moving anything forward. They finished in 13 days. That ratio - 18x - is the number everyone is sharing, and it deserves a closer look than the headline usually gets.
The migration did not happen because an AI wrote better code than an engineer. It happened because the team built a rule-based framework using Claude - markdown files combined with reference implementations - to standardize the AI-automated migration , then fed each round of PR feedback back into the rule set so accuracy kept improving, let autonomous LLM loops of building, fixing, and validating run without manual intervention, and parallelized migrations across isolated environments . The people did not disappear. The 18x speedup was not the model working alone, but engineers scoping the task, writing the rules, building reference implementations, and reviewing the output while the agent handled the repetitive cycle. Salesforce's own account credits structure and human oversight as much as the model. The people did not disappear from the process; their work shifted from writing every line by hand to defining the problem precisely and supervising an agent that executed it.
That distinction matters, and it is easy to lose in the headline.
What the Salesforce numbers actually show about agentic coding workflows
Salesforce's head of engineering published a post saying the company moved its entire development organization to agentic workflows. They rolled out Anthropic's Claude Code across the whole company as the main AI agent and gave every developer unlimited tokens to use it. The org-wide results: in April 2026, work items completed per developer are up 50.8% compared with April 2025; PRs merged per developer are up 79%; and using a machine learning-based Effective Output score, output has grown 151.3% year over year.
None of these numbers can be independently verified. That is worth stating plainly - this is a company reporting on itself. But the methodology is specific enough that it is hard to dismiss outright, and the 13-day migration case study has the kind of checkable structure (33 endpoints, 5 PRs, one PR with 21 endpoints at full test coverage) that makes fabrication implausible.
The productivity and quality tradeoff most people assume would appear here did not. Despite the surge in pull requests, incidents dropped five percent.
The more interesting question is what changed structurally. One of the most interesting developments has been watching engineers become builders of their own agentic workflows, not just users of tools. Claude Code skills - packaged, reusable capabilities that encode team context, naming conventions, and workflow patterns - have become a new form of engineering artifact. Teams are building them, sharing them, and compounding on each other's work. That compounding is the signal. The 13-day migration was not a one-off trick. It was the result of an organization building institutional knowledge about how to structure problems for an agent, and encoding that knowledge into reusable files rather than in individual engineers' heads.
What Anthropic's own session data adds to the picture
The Salesforce post has a useful counterpart. Anthropic published research this week drawn from a privacy-preserving analysis of roughly 400,000 interactive sessions from about 235,000 people between October 2025 and April 2026. It describes what actually happens inside those sessions, and the findings complicate the "the AI does the work" narrative in a useful way.
In a typical session, people make most of the planning decisions - what to do - and Claude makes most of the execution decisions - how to do it. The greater domain expertise a person brings to a session, the more work Claude does per instruction. That is a clean formulation of something practitioners feel but rarely articulate: the tool is not a replacement for knowing your domain, it is a multiplier of it. An engineer who understands the problem deeply gives better instructions and catches errors faster. The agent does more because the human is better, not instead of it.
The composition of the work done with Claude Code changed substantially between October 2025 and April 2026. The clearest change is that the share of sessions spent fixing broken code fell from 33% to 19%. In its place, there was a greater share of the work that surrounds code. Operating software grew from 14% to 21% of sessions. Writing and data analysis roughly doubled, from about 10% to 20% of sessions. Less debugging, more building and operating. By the measure of estimated freelance marketplace value, the estimated value of the average session rose by 27% between October and April. The work being delegated to agents is getting more expensive, not cheaper.
The part of the Salesforce post nobody is quoting
The productivity claims got the coverage. The harder problems in the post did not. Context management in long, complex agentic sessions remains a craft that engineers are actively learning. The quality of CLAUDE.md files - the persistent context configurations that orient Claude to your codebase, conventions, and constraints - varies widely across teams, and that variance matters a lot for output quality.
This is the thing most teams will actually run into. The 13-day migration was not the default output of pointing Claude Code at 33 endpoints. It was the output of a team that had already put in the work to define what good looks like for their codebase, and encoded it in a form the agent could use. The most important skill today, in the words of Salesforce's engineering head, is knowing how to structure problems for an agentic system, when to delegate versus stay in the loop, and how to build reusable patterns your team can compound on.
That is a different skill than writing code. It is closer to technical program management - breaking work into verifiable chunks, writing precise specifications, knowing when the agent has gone off-track before it has gone too far. When agents handle more of the execution layer, the question of how junior engineers grow into senior engineers becomes real, if AI is absorbing much of the entry-level work. Salesforce flags this as an unsolved problem. That is the right frame.
Agentic coding has taken off. The share of GitHub projects with coding agent activity has more than doubled since late 2025, and Claude Code users now spend an average of 20 hours per week using the tool. At that scale of adoption, the teams who figure out the craft layer - the CLAUDE.md files, the rule frameworks, the compounding skill libraries - will pull ahead of the teams who are simply "using AI" because it shipped with their editor.
What the 18x number actually asks of teams
The honest version of the Salesforce story is not that 231 days became 13 because AI is fast. It is that a team invested in understanding how to work with agents - built the scaffolding, wrote the rules, ran the feedback loops - and then the speed showed up. The roles Salesforce describes still require people to define the problem, set the rules, and verify the result - judgment that AI does not replace. The shift is in what an engineer spends time on, and how many are needed for a given volume of work.
That is the question worth sitting with: not "can our team use Claude Code" but "have we done the work to tell an agent what good looks like for our codebase?" The 18x is downstream of that answer, not a shortcut to it.