Few matters spark as a lot debate proper now because the “agentic shift” in coding. As a substitute of writing code line by line, builders orchestrate software program creation by AI brokers.
Salesforce is now placing its personal numbers behind that shift. In a submit by Srinivas Tallapragada, Salesforce’s head of engineering, the corporate says it has moved its complete improvement group to agentic workflows. They rolled out Anthropic’s Claude Code throughout the entire firm as the primary AI agent and gave each developer limitless tokens to make use of it.
For April 2026, Salesforce reviews a pointy effectivity soar in comparison with the identical month final yr. Accomplished work gadgets per developer rose 50.8 p.c. Merged pull requests per developer climbed 79 p.c.
An ML-based “Efficient Output Rating” designed to measure the precise worth of shipped code improved by 151.3 p.c. None of those numbers might be independently verified.
Extra output, fewer incidents
The apparent query, whether or not high quality suffers at this tempo, Tallapragada solutions by pointing to the corporate’s personal monitoring platform, Engineering 360. Regardless of the surge in pull requests, incidents dropped 5 p.c. Security guardrails and high quality requirements are baked into the agentic workflow, he says.
“When agentic instruments get utilized correctly, high quality does not endure from pace. It advantages from it,” Tallapragada writes. Salesforce does not again this declare with exterior audits or impartial measurements.
Engineers are actually constructing their very own agentic workflows relatively than simply utilizing off-the-shelf instruments, in accordance with Tallapragada. So-called Claude Code skills, reusable capabilities that encode crew context, naming conventions, and workflow patterns, have turn out to be a brand new type of engineering artifact. Salesforce additionally constructed a curated library known as “AI Knowledgeable Suite” and “Salesforce Basis Plugins” that serves as a shared basis for all builders.
Sub-agents and agent groups, specialised AI brokers that deal with parallel workstreams inside a bigger activity, are altering how complicated work will get damaged down. Builders not bounce between 5 methods. They describe the specified final result, and coordinated brokers deal with the person steps.
API migration in 13 days as a substitute of 231
As a concrete instance, Tallapragada factors to migrating 33 API endpoints to a brand new cloud-native structure. The standard method would have taken about 231 person-days, the corporate estimates. Utilizing a rule-based framework constructed on Claude with Markdown information and reference implementations, the migration was finished in 13 days; 18 occasions quicker.
Every spherical of PR suggestions was fed again into the rule set, so accuracy saved bettering. Autonomous LLM loops of constructing, fixing, and validating ran with out guide intervention. Migrations had been parallelized throughout remoted environments. The end result: 5 pull requests, with the most important single PR delivering 21 endpoints with full check protection.
“Crucial talent right now is realizing find out how to construction issues for an agentic system, when to delegate versus keep within the loop, and find out how to construct reusable patterns your crew can compound on,” Tallapragada writes.
Safety, junior expertise, and crew construction stay unsolved
Tallapragada is upfront a few vary of unsolved issues, calling them “genuinely laborious.” Context administration in lengthy agentic classes is a talent engineers nonetheless must study. The standard of CLAUDE.md information—persistent context configs that align Claude with a codebase—varies extensively between groups and has a huge impact on output high quality. Safety wants a rethink too. When brokers act on methods relatively than simply making strategies, the blast radius of a misconfigured device will get a lot bigger.
Then there’s the expertise pipeline query. “When brokers deal with extra of the execution layer, how do junior engineers develop into senior engineers if AI is absorbing a lot of the entry-level work? What’s the function of a designer or product supervisor on this new world?” Tallapragada writes. Salesforce is experimenting with one-person or three-person items as a substitute of conventional Scrum groups. It does not have clear solutions but.
Productiveness leap or tech debt on autopilot?
A sharply completely different take got here a couple of days in the past from well-known programmer and hacker George Hotz. Utilizing AI brokers in software program improvement will likely be one of many trade’s most costly errors, he argues.
LLMs are “subtle statistical fashions” that “mimic the distribution of programming” however can by no means really program, Hotz says. Massive organizations are particularly in danger as a result of weaker builders cannot spot defective output.
Even Andrej Karpathy, who now counts himself amongst agentic coding’s supporters, has flagged high quality issues. Agent-generated code is “not like tremendous superb code essentially on a regular basis,” he mentioned, calling it “bloaty, there’s a number of copy paste, there’s awkward abstractions which might be brittle, and like, it really works, nevertheless it’s simply actually gross.” In contrast to Hotz, although, Karpathy remains to be offered on the brand new method and lately joined Anthropic.
A broader debate concerning the rising prices of AI relative to its advantages is heating up too, alongside questions on what the fashions really ship in day-to-day work.
AI Information With out the Hype – Curated by People
Subscribe to THE DECODER for ad-free studying, a weekly AI e-newsletter, our unique “AI Radar” frontier report six occasions a yr, full archive entry, and entry to our remark part.
Subscribe now