A $3,000-a-Month AI Workflow Still Needs Human Taste
A Salesforce research engineer runs five to six coding agents and spends about $3,000 a month on tokens, yet hand-wrote his most successful tool because agents produce slop without human taste.
Ben Stolovitz, a research engineer at Salesforce AI Research, reports that agentic coding is now his primary workflow: he manages five to six coding agents in parallel and spends about $3,000 a month on tokens. The agents handle tests and pull requests, but he still validates their work rather than trusting them unsupervised.
He also uses LLMs for research, learning, summarization, and voice dictation, and has drafted a personal AI policy. He remains skeptical of AI writing and image generation, and says agents still need human taste to produce good work.
Key takeaways
- Each agent works in its own Git worktree, writing code, running tests, and creating or reviewing pull requests; prototypes arrive in roughly 20 minutes.
- Human taste remains decisive. Left alone, agents produce over-architected, unreadable slop; Stolovitz hand-wrote a tool that replaced about 4,000 lines of vibe-code with a few hundred, and he reviews architecture rather than line-by-line.
- Research and learning benefit from LLMs, especially Google's AI summaries, pub facts, and Socratic tutoring, but novel research suffers: agents led him astray on turn-detection architectures and could not annotate training data correctly. Human experts stayed more reliable.
- Writing is still off-limits for creation. He does not use AI to write from scratch, finds AI prose plodding and repetitive, and rewrites agent-generated documentation; AI editing is mediocre and sycophantic.
- Art and music remain mixed. He stopped generating images, enjoys some AI music but dislikes its commercialization, and sees AI video as mostly slop. AI-generated art has become ubiquitous in ads, event posters, restaurant menus and store logos.
What people think
Analysis details
16 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 16 of 22 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
AI music's human-expression test
AI music provoked the sharpest split. One position holds that it is hollow and pointless because it lacks human expression, and a waste of compute that enriches platform owners; the other holds that human intent remains, that AI can pair off-beat ideas with polished execution, and that it beats mass-market muzak. The disagreement turns on whether novelty and execution can substitute for human expression.
Domain modeling remains the hard part
The claim that writing code is solved drew pushback: syntax and CRUD are largely automated, but modeling a domain into data structures remains hard. One view holds that 85-90% of developers struggle mainly with syntax and database rows, so AI covers most software work; another holds that encoding real-world concepts into types is the actual difficulty, and that non-technical users may build adaptable software because they can articulate their wants.
Agent code review still needs human eyes
Agents bake incorrect assumptions into code and miss them in tests, making code reading and reasoning-log inspection necessary. A prompt asking for issues that are 'both critical and related' was interpreted as exactly two issues, though another view treats the example as user error rather than model failure. Oversight also carries cost: more than $3,000 a month in tokens, by one self-reported account.
Counterexamples to solved coding
A Gemini outage, a hallucinated package name, and a site that breaks without storage access were cited as evidence that bug-free software is not the norm. The examples question the claim that coding is solved, though they are anecdotal rather than a systematic measurement.
Training-data degradation and RL
The concern that model quality will degrade as more people rely on LLM outputs was challenged by the claim that reinforcement learning does not depend on training data and that the predicted 'mathocalypse' did not occur because better data was found. That remains a contested forecast rather than a settled outcome.