Build Salesforce features with eleven subagents that check each other's work

A Salesforce builder split Claude Code into eleven subagents that research, design, build and test as real users. 366 runs in two and a half weeks cost about $1,776, and nothing has reached production yet.

Done withClaude CodeinSalesforce· Nov 18, 2026

What they did

Pranav Nagrecha uses Claude Code on a laptop to build Salesforce flows, pages and rules for business teams. After a first week where one session built a feature, tested it and marked it passed, the author split the work into eleven subagents, each a text file with one job: a researcher, an analyst, an architect, a test designer, a builder, structural QA, runtime QA that logs in through a browser as each kind of real user, a verifier, a documenter, and a guide writer with a separate reviewer on a different model. Skills call the agents in order, a main session reads their reports and decides the next step, and the author approves before every build. A script checks every command that would change Salesforce and refuses anything not aimed at the project's own developer sandbox.

How it went

The split caught real problems. A flow the single agent called done had 137 elements; agreeing the size up front brought the redesign to 57. The architect sent back 60 of 126 test rows on first review, and the guide reviewer found 15 errors in the first user guide. Over two and a half weeks 366 agent runs cost about $1,776, with the main session taking 38% of that. Even so, the author's own pass over one finished project found eight faults no agent flagged. Two projects sit in a shared test sandbox and none has gone to production.

Worth knowing

Tests run only as an admin miss what real users see, because admins can skip validation rules; the author removes that permission and logs in as each user type. The sandbox guard lives in the project folder, so the agent could edit it before a run, which is not fixed yet. A flow can deploy and pass checks while carrying a data-type error only visible when opened in Flow Builder.

Try it

With Claude Code· 4 parts to fill in

Set up a set of Claude Code subagents for building on Salesforce in [project folder]. Create separate agents for requirements, design, test design, building, structural review and runtime testing, and a skill that runs them in that order. The test designer writes tests from [requirements file] before anything is built and never reads the build. The builder only deploys to [developer sandbox alias] and stops to ask me when Salesforce refuses something instead of building a workaround. Runtime QA runs every test as an admin without the bypass-validation permission and then as each of [user profiles]. Show me the list of what will be built and wait for my yes before every build.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related