How I'm fighting the AI brain fry
Growing reliance on AI coding agents is leaving many developers feeling cognitively overloaded, as they juggle multiple long‑running sessions, constant context switching, and large AI-generated code changes to review. Commenters propose various coping strategies—from kanban-style queues, stricter task scoping, and limiting parallel work to using AI mainly for explanation or review rather than generation—while others argue for simply accepting more idle time and breaks. A strong undercurrent of skepticism targets AI-written prose and “slop” content, reflecting broader worries about atrophy of deep understanding, homogenized writing styles, and a culture that prioritizes speed and automation over thoughtful work.
Reality of “AI Brain Fry”
- Many commenters say cognitive exhaustion from AI-heavy workflows is real, especially for experienced programmers who now feel unprepared for traditional interviews or deep work.
- Others argue the panic is overstated: problem‑solving abilities can be exercised in many ways, and what mainly atrophies is coding fluency, not general intellect.
- Several distinguish raw cognition from higher‑level habits of thinking and understanding, which they see as more at risk when outsourcing too much to agents.
Multitasking, Context Switching, and Productivity
- Core tension: run many agent sessions in parallel vs. focus on one task.
- Some run 10–30 concurrent agents, likening it to managing a team, or even playing many games of chess at once; they say attention becomes about quick, local decisions, not holding all context.
- Others cap themselves at 1–3 active tasks, reporting better focus, less fatigue, and often equal or higher effective throughput.
- Multiple people note that context switching feels productive but often slows net delivery, especially due to large PRs and review burdens.
Workflows, Tools, and “Kanban for Agents”
- Common pattern: treat agents like team members managed via tickets, kanban boards, or git‑resident issue files.
- People build bespoke tools (TUI/GUI boards, local issue trackers, “your turn / AI’s turn” queues) or adapt GitHub/Jira/markdown/org-mode to track sessions, decisions, and status.
- Several emphasize small, well‑scoped tasks, explicit specs, and front‑loaded “grilling” (Q&A) to reduce mid‑task interruptions.
AI Writing Style and Backlash
- Strong negative reaction to the blog’s partially AI‑generated prose; many say they can “smell” the model style and find it verbose, clichéd, and dehumanizing.
- Some argue AI‑written blog posts signal low effort and erode trust; a few say they now stop reading as soon as they detect “AI slop.”
- Others are indifferent: if the text communicates the idea, they don’t care whether a model drafted it, especially for non‑literary writing.
Code Quality, Speed, and Human Understanding
- LLMs greatly accelerate drafting code, tests, and refactors, but reviewers report brittle abstractions and tech‑debt seams that still require substantial human pruning.
- A recurring worry: you don’t internalize AI‑written code like your own, weakening long‑term understanding.
- Fast models reduce the urge to parallelize, but external bottlenecks (builds, tests, CI, tools) keep “agent wait time” significant.
Human Limits and Cultural Questions
- Multiple comments suggest using “agent thinking time” as a break—reading, walking, music, reflection—rather than spinning up more work.
- Some call for resisting the industry’s speed‑at‑all‑costs culture, emphasizing deep focus, reading books, and doing less but better.