Your Attention Is the Bottleneck, Not Your Agents — Zack Proser, WorkOS (Category AI4SDLC)
I saw it the other day. report Zack Proser from WorkOS, which talks about AI burnout from working with agents. I think you know what it feels like to work with Claude or Codex, when it seems like more has been done than before, but by the middle of the day, you’ll be out. I really started feeling that when I was running agents all day. You need to formulate the problem, hold the context, choose the boundaries, check the result, then explain again where the agent missed. Technically, the speed is higher. But the feeling of flow, which used to appear when you sit down to write a complex document or code a non-trivial feature, almost no. Instead of flow, it’s a manager’s day: lots of context switching, making decisions, and controlling what agents do.
In fact, Zack’s main thesis from the presentation is that agents are no longer a bottleneck, and human attention remains a finite resource. If you give agents context, test criteria and tools, they can spin almost indefinitely, he said. But a person must still understand that the task is solved, the quality is sufficient, and the solution coincides with the product and engineering reality.
The report has a good example from his work at WorkOS’ Applied AI team. Slack-bot for internal blog posts broke sentence case: bot spoiled acronyms like SSO. Instead of jumping between Slack, Linear, the terminal and the browser, Proser gave Claude Code access to Slack and Linear via MCP, asked to fix the bug and check the result in the same loop. The important detail here is not "the agent wrote the code", but "the agent reached the completed verification loop".
Next Zack suggests to optimize their work by layers
1️⃣ Signal layer If Slack or Linear is your main source of switches, you don’t have to open them yourself. An agent can read mentions, DMs, tickets, deduplication queries, and pull out only what really needs attention. This does not negate communication, but protects focus from random noise. 2️⃣ Voice-first flow Zach says he regularly voices around 184 This helps to run several working branches in Claude Code, Cursor or Codex faster. But for me, the caveat is important: voice does not remove the need to think. It reduces the friction between intention and task setting. If the intent is cloudy, the agent will build the wrong thing faster. 3Remote control and shower principle Some of the good decisions don’t come in the IDE, but on a walk or in the shower when the brain comes out of the rigid focus mode. Before, moving away from the table meant stopping work. Proser shows a different mode: in the morning load agents with tasks, give them context and checks, then walk away from the table and direct them from the phone when a clearer thought appears. I'm not sure that's a good idea for everyone, but it's a valuable idea: productivity doesn't equal sitting in front of a screen. 4Safety and verification gates Zach formulates this almost as an engineering rule: speed requires safety. Minimum gate - lint, build, unit tests. The next is a browser check: the agent passes the script himself and makes sure that, for example, the login is not broken. Even higher is the critic pass, when an individual agent checks the result against a set of rules. Here, agent development ceases to be a prompting and becomes a process.
I also liked the idea of a weekly pass on the history of agency sessions. With Claude Code, conversations are stored locally in JSONL, and Proser suggests regularly giving them to the agent with the question: where we spent unnecessary thinking tokens, where we removed ambiguity, what skills, MCP servers or templates were missing. We usually throw out the context of post-mortem work, although that is where the map of recurring attention loss lies.
There is also a funny but accurate episode about Oura Ring via MCP: an agent may remind you that you did not sleep well and you should not take the second half of the plan. Against the background of “burnout turbo”, as Proser describes the mindless acceleration of LLMs, even such a pause becomes part of the operating model.
The Q&A report had an important question about educating people. Proser’s answer would be: Don’t delegate to AI what you don’t yet understand. If you're learning, you still have to go through pain with your hands, build mental models, and learn to catch bad decisions. AI can speed up learning and explain gaps, but it doesn’t have to steal the underlying engineering muscle.
#AI #AI4SDLC #Engineering #Software #Management #DevEx #Process