TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
In an October 2026 personal essay, Salesforce AI Research engineer Ben Stolovitz describes a year of major changes in how he uses AI, especially for software development. He says he now manages several coding agents at once, while stressing that human judgment remains necessary and that agent work can be tiring and diminish the pleasure of writing code.
Ben Stolovitz, a research engineer at Salesforce AI Research, says his use of AI changed substantially over the year leading to October 2026: instead of writing much of his code himself, he now directs several coding agents to implement features, run tests and prepare code reviews. In a personal account, he also describes drafting rules for his own AI use and says the agents still need human judgment, while their use can be mentally exhausting.
Stolovitz says he typically manages about five or six agents, using tools including Claude Code, OpenCode, Codex and Pi. He gives each a separate copy of a codebase and assigns work such as implementing a feature or reproducing a bug. According to his account, agents can then write code, run tests and create and review pull requests while he works on other tasks.
He describes the speed as useful for prototyping and experimentation, including a change to the open-source Mumble client intended to improve noise cancellation. Stolovitz says he uses a beta version of that change nearly every day. He also reports using agents to test audio-related code on macOS and iOS and to diagnose and fix bugs. These are examples from his own work, not independently evaluated results.
His account covers other uses too: research and search, summarization, transcription, voice, writing, art and music, and video. He says he spends roughly $3,000 a month on tokens, but does not specify the precise period or accounting method behind that estimate. He describes the account as personal opinion and speculation, rather than a formal study of AI use.
Coding Shifts From Writing to Supervision
Stolovitz’s account offers a practitioner’s view of a workplace change that is easy to miss when AI coding is described only as autocomplete: his day is increasingly spent assigning and checking tasks rather than writing code line by line. He says agents make it easier to try ideas quickly and propose large changes, but his description also shows that delegation does not remove engineering work. Reviewing output, directing agents and deciding whether a design is sound remain part of the job.
He warns that agents can produce poorly functioning, overbuilt or hard-to-read code if left to work without guidance. In one case, he says he replaced about 4,000 lines of code produced with AI assistance with a few hundred lines in a tool he wrote himself; he found the replacement more flexible and easier to test. The example illustrates his view that speed and volume are not the same as quality.
The account also raises a human cost. Managing multiple agent conversations can feel overstimulating and exhausting, Stolovitz says, and handing off code can take away the craft and satisfaction some programmers find in designing and building software. His experience is one person’s, but it gives readers a concrete picture of both the productivity claims and the trade-offs involved.
Products Worth Considering
Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Bluetooth Noise-Cancelling Over-Ear Headphones, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Hazelnut Taupe
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
As an affiliate, we earn on qualifying purchases.
A Year After Copilot Autocomplete
Stolovitz compares his current routine with an account he wrote in October 2025. At that time, he highlighted GitHub Copilot’s code autocomplete. Twelve months later, he says autocomplete remains useful but is no longer central to his work. He attributes the shift to agentic coding: tools that can take on broader tasks, rather than simply suggest the next lines of code.
His professional setting also changed during that period. Stolovitz says he moved from being an engineer on Microsoft’s Windows AI team to a research engineer at Salesforce AI Research, where he spends more time supporting scientists and building rapid prototypes. That background helps explain the kinds of work he describes, but does not establish that his routine is typical of engineers generally.
Alongside describing tools, he says he has begun drafting a personal AI policy to guide everyday decisions as AI spreads into work and personal life. He frames the document as an evolving draft, not a finished set of rules, and says he wants to distinguish uses that help him make things with other people from those that may undermine that aim.
How Typical Is This Workflow?
The account does not establish how common Stolovitz’s approach is among software engineers, how often the agents’ work requires substantial correction, or how his reported pace compares with a team working without them. It is a personal report, not a controlled assessment of productivity, code quality or cost.
His estimate of about $3,000 a month in tokens is not accompanied by a breakdown of tools, usage or the period measured. The article also leaves open how durable the reported benefits will be and whether coding agents will become better at producing readable, maintainable systems without close supervision. Stolovitz says the duration of the human advantage is uncertain.
A Personal Policy Still in Draft
Stolovitz says he plans to keep developing his personal AI policy and adapt it as the tools and his experience change. He also says he is considering a separate AI policy for his website, but has not finalized it. No publication date or final policy text is provided in the account.
For his own work, the next question is how he will balance the speed of delegating tasks with review, focused attention and the satisfaction of making things by hand. His essay offers a snapshot rather than a forecast: it records what changed for him over one year and leaves open how quickly the tools, norms and costs will continue to change.
Key Questions
What changed in Stolovitz’s use of AI?
He says he moved from relying mainly on code autocomplete to managing several coding agents that can implement tasks, run tests and prepare pull requests.
Which coding agents does he mention?
He names Claude Code, OpenCode, Codex and Pi, and says he typically manages about five or six agents at a time.
Does Stolovitz say AI has solved software engineering?
No. He says agents still need human judgment and can produce code that is poorly designed, difficult to read or hard to maintain.
How much does he say he spends on AI?
He estimates spending roughly $3,000 a month in tokens. He does not provide a detailed breakdown or a precise accounting method.
Is this evidence that all engineers are using AI this way?
No. The article is Stolovitz’s personal account and does not measure how widespread this workflow is or compare its results with other approaches.
Source: ben.stolovitz.com
