When AI agents make code cheap, the scarce work becomes the checking that decides what reaches production. Four accounts cover who may build first, what evidence a change must carry, how much testing a system can absorb and how much one person can review.
Will Larson (22 Nov · 01), an engineering executive at Imprint, a company that runs branded credit card programmes, describes how his teams now start new work. Under what he calls the Explorer Pattern, a group of at most five people builds a working version end to end before other teams are consulted. The group then reconciles the details with those teams and accepts the rework that follows. One engineering leader built support for Plaid, a service that links bank accounts to other apps, as a low-cost side project, and it shipped without ever appearing on a roadmap. Larson sets conditions: a named sponsor for any project past about three weeks, progress posted in public, and a feature flag that only staff can see.
Kamal Devarakonda (22 Nov · 02), a senior director of software engineering at Salesforce, leads a team that reviewed 15 to 20 production changes a week across 18 Salesforce products. Each change needed approvals from quality assurance, two developers, an architect and a cloud leader, and gathering that evidence took about a day. The team built an AI agent called ROAM to do the review, with ordinary code checking each approval against the records. The AI model only explains the result, and when an approval cannot be confirmed the request goes to a person, about 8% of the time. Over about 550 reviews in 10 weeks the agent made no incorrect approvals, and release engineering made it mandatory.
A developer who writes as tettuan (22 Nov · 03) published an account in Japanese of building a market analysis app with four Claude Code AI agents writing all the code. In October the project stalled for four weeks. About 150 to 200 pull requests a week queued for one test machine per operating system, and jobs waited more than 100 hours a week. The author kept the number of test machines fixed and changed how agent work reached them: one integration branch, bundled issues and one full test run per batch. Closed issues rose from about 30 to more than 100 a week.
Pavel Jiranek (22 Nov · 04), who writes the newsletter The Hybrid Leader, describes supervising AI coding agents in his own software work. During release preparation he may have up to nineteen agents fixing bugs, testing, documenting and reviewing. He finds his limit when he returns to a task and its context no longer comes easily. He cites a survey of 203 users of AI coding agents: 83% reported a lighter workload, while 101 also rated at least five supervision tasks as very demanding. His advice is to count the outcomes still awaiting a person's judgement, and for teams to plan review time as deliberately as coding time.
In all four accounts, the check before production limits the pace more than producing the work does. Imprint moves that check until after a small group has built something real. Salesforce puts it in code that reads the records, and the solo developer runs it on fewer, larger batches. Jiranek treats one person's capacity to verify work as the limit on how many AI agents to run. On who decides which prototypes deserve engineering's review time, Larson's answer is a named sponsor. Salesforce's answer is evidence that code can confirm before a person looks. This comparison is an inference and is not either author's claim.
Each account comes from a setting that suits it. The four settings are a company where one executive sets the process, an internal release flow with structured records, a one-person project and one developer's own routine. A larger organisation with shared systems may find that reconciling a finished prototype with other teams costs more than agreeing the design first. Jiranek himself notes that the studies he cites are early and set no safe number of AI agents.
Will Larson of Imprint describes the Explorer Pattern, in which up to five people build a working feature before consulting other teams, because cheap code makes rework affordable.
Salesforce's release team built an AI agent that reviews production changes in seconds, while ordinary code, not the AI model, decides whether each required approval exists.
A developer running four Claude Code AI agents tripled the issues closed each week by batching their work through one integration branch after testing capacity stalled the project.
Pavel Jiranek, who runs up to nineteen AI coding agents during a release, argues that the work awaiting a person's review sets how many agents that person can manage.



