AI Boosts Output but Destroys Developer Focus. What Now?

The claim. AI tools like coding agents and second-brain systems dramatically increase individual output, but they exact a hidden cognitive cost. The speaker argues that running multiple agent sessions in parallel forces developers into constant context-switching, which fragments attention, blocks flow state, and leaves them exhausted by end of day. He calls this the "fried brain effect," citing a tweet that resonated widely: "AI was supposed to make us work less, but I've never been this tired in my life."

Why it happens. Developers now juggle several agent conversations simultaneously, each requiring supervision and intermittent input. The speaker describes how he sends a request to one coding agent, then switches to another while waiting, and another after that. The guilt of idle time—knowing you could spin up yet another agent—creates a treadmill of parallel sessions. Humans cannot parallelize cognitive attention the way AI systems do, so each switch costs focus and recovery time.

The in-between. The speaker frames this as a temporary phase. Current LLMs and agent harnesses are not autonomous enough to let projects run unattended for hours. Most tasks require "babysitting," which means developers must monitor, approve, and course-correct frequently. Over time, agents will improve, and the need to babysit will shrink. But right now, every improvement in agent capability tempts users to add another session rather than reduce the burden.

The proposed solution. Rather than spinning up more parallel sessions, the speaker advocates investing time in building a more autonomous harness—what he calls a "software factory." The idea is to codify workflow, validation rules, and approval gates so the agent can plan, implement, and verify work with minimal human intervention. This sacrifices short-term throughput for long-term cognitive relief: fewer switches, batch reviews, and one primary planning session per day.

Tradeoffs and open questions. Building a custom harness requires significant up-front effort and is not accessible to everyone. The speaker acknowledges that some developers will prefer to wait for out-of-the-box solutions. Even then, a commenter in his community pointed out that once one factory works, teams might just run multiple factories in parallel—potentially recreating the same problem at a higher abstraction level. The speaker admits this is possible but hopes a single primary session can act as the orchestrator.

The bigger picture. The video is part observation, part vendor positioning (Revolut Business sponsor), and part self-promotion for the speaker's software factory project. The core insight—that AI's productivity gains come with attention fragmentation—is real and broadly experienced. However, the evidence is anecdotal, and the proposed fix is speculative and unproven at scale. Engineering leaders should take the problem seriously but treat the solution as one person's experiment, not a proven pattern.