AI agents can scale their work faster than a developer can maintain attention and judge quality.
2
Signal layers, voice input, remote control, and verification gates let developers direct agents without staying at the desk all day.
3
Reviewing conversation history can reveal repeated struggles and produce skills that make future work more efficient.
Summary
Zack Proser argues that AI coding tools have shifted the limiting factor from agent capacity to human attention. He describes a Slack bug fix where Claude Code read the request, changed the code, triggered the relevant Slack workflow, and verified the result. His proposed workflow moves context gathering to agents, uses voice to dispatch work quickly, and lets developers monitor sessions remotely from a phone. Verification gates range from lint, builds, and unit tests to browser checks and another agent's review. Proser also recommends analyzing Claude Code's JSONL history to find repeated friction and create new skills. He is careful about the limits of this approach. Developers still need judgment, quality control, and hands-on learning, especially early in their careers. An Oura Ring integration adds a blunt reminder that sleep affects the work system too.
Proser says agents can handle a large number of tasks when given enough context, tools, and verification criteria, but the developer cannot inspect all of that work indefinitely and return for another full day without burning out. Agents can scale through more parallel sessions, while human attention still degrades under load. He cites Simon Willison's experience of starting four parallel agents and being exhausted by 11 a.m. The practical problem is finding a personal limit rather than assuming that more agent capacity should automatically mean more output.
A signal layer keeps Slack and Linear from stealing focus
Proser's first layer puts an agent between the developer and incoming work. Instead of opening Slack and getting pulled into unrelated threads, Claude Code reads Slack on a loop, checks mentions and direct messages, and identifies high-priority requests. It also has access to Linear through MCP so it can connect asks to the actual tickets and remove duplicates. This gives the developer a filtered view of work while reducing the context switching caused by constantly checking messages and project updates.
Proser has used voice-first coding for about a year and a half. He says he regularly reaches 184 words per minute, compared with a previous typing speed of about 90 words per minute. The value is more than entering one prompt faster. He can speak across several Cursor windows, Codex, and Claude sessions while another developer might still be typing the first request. As he becomes more comfortable with voice, he can start work across multiple agents and spend less time physically sitting at the desk.
Remote control lets developers leave the desk without stopping work
Proser connects the shower principle to remote agent control. Focus at an IDE helps with execution but can also create blind spots, while walking, playing with a dog, or taking a shower can produce a fuller solution. With Claude Code remote control, a session keeps running on the development machine while Proser checks it from Claude on his phone over another network, including LTE miles away. He can send a new design idea immediately, review pull requests from outside, and return to work that has already incorporated the message.
Proser says speed requires safety and describes several verification levels. The first runs lint, builds, and unit tests through hooks after code changes. The next asks the agent to use a browser, click through a flow, and check that something such as login still works. A further level uses another agent to compare the result with a stated constitution or set of requirements, then send feedback when the work falls short. These checks allow agents to continue while the developer is away, although the developer remains responsible for review and shipping.
Weekly history reviews can turn repeated friction into skills
Claude Code stores conversations locally in JSONL files, which gives Proser a record of how work was done. He recommends a scheduled daily or weekly pass that looks for places where a task required many tokens, repeated clarification, or prolonged back-and-forth. The agent can then identify missing MCP servers, tools, or skills and suggest ways to tighten the loop. Proser says Claude Code can build, evaluate, and improve skills from natural-language instructions. Hooks can also extract the useful parts of each session into a smaller Markdown or other data store before analysis.
Personal health data can become part of the working loop
Proser connected his Oura Ring to Claude through MCP. During an argument about a project, Claude has told him that he did not sleep and suggested doing only the first part of the work. Proser admits that he sometimes ignores the recommendation, but says the exchange at least makes him consider taking a break. He wants the system to account for more than tickets, conversations, and coding skills. Sleep and the times when he can focus are also inputs to how much work he should direct.
Early-career developers still need to build skills by hand
In response to a question about skill development, Proser recommends not using AI for work the developer does not understand yet. He uses AI to move faster on TypeScript, RAG systems, and AWS deployments because he learned those areas through years of difficult work and can spot bad suggestions. People early in their careers should still go deep, code some things by hand, and learn what makes them painful. Once they have confidence in a skill, they can delegate more. He also suggests asking an AI system to test their understanding and expose gaps.
"Our attention is still in meatspace, if you will, and it still degrades under load."03:41
Who should watch
You are running several coding agents and ending the day exhausted despite shipping more work.
You want to monitor long-running agent sessions away from your desk without giving up review and control.
You are early in your career and need a way to use AI while still learning systems deeply by hand.
