Simon Willison, on Lenny’s Podcast in April 2026: fire up four agents on four problems and by about 11 AM he is wiped for the day. The agents were not the tiring part. The watching was.
High AI oversight - watching, checking, and steering tools and agents - tracks more mental load than raw AI use. In BCG’s March 2026 survey of 1,488 full-time US workers (published in HBR as “AI brain fry”), fourteen percent of AI users reported brain fry. People with high vs low oversight also reported +14% mental effort, +12% mental fatigue, and +19% information overload (same survey). Self-reported productivity rose through three simultaneous AI tools, then dipped. Microsoft Research’s ParallelPilot work (formative N=14, lab N=16) finds the same bottleneck in coding: the tax is coordinating and monitoring concurrent sessions. ParallelPilot’s five practices (Plan, Isolate, Log, Observe, Triage - PILOT) cut that tax in the lab and raised short-task throughput 63%.
Definitions
AI brain fry - mental fatigue from excessive use or oversight of AI beyond cognitive capacity. People described buzzing or fog, hard focus, slower decisions, headaches. Not the same thing as chronic burnout.
Oversight - how much the tool needs your direct monitoring (reading and judging outputs, answering agents, steering fleets) rather than quietly replacing routine toil.
Parallel AI coding - running multiple independent coding-agent sessions at once (separate CLIs, worktrees, or cloud runs) and supervising them yourself.
PILOT - five supervision practices from Microsoft Research’s ParallelPilot study: Plan, Isolate, Log, Observe, Triage.
Attention residue - leftover activation from an unfinished or interrupted task that steals capacity from the next one (Sophie Leroy).
Oversight is the load, not “using AI”
BCG’s authors split two patterns. When AI replaces repetitive tasks, workers reported about 15% lower burnout. When the job meant high oversight of AI, mental effort, fatigue, and information overload went up. Brain fry clustered with early adopters and people running many tools or agents - the cohort companies praise for leaning in.
Quit intent among people who reported brain fry was 34% vs 25% without it. Error-frequency scores were higher too (Register summary: about 39% higher among those who felt they were using too much AI). This is survey self-report, not a clinical diagnosis and not a peer-reviewed instrument with published p-values. The direction still matches what parallel-agent users already feel.
Willison’s 11 AM is the lived version of those numbers: four agents, four problems, and the day’s mental budget is gone before lunch.
How the workday densifies
Berkeley Haas researchers Aruna Ranganathan and Xingqi Maggie Ye spent about eight months (April-December 2025) inside a ~200-person US tech company with broad GenAI access - on-site observation plus more than 40 interviews. Nobody mandated use. People still intensified work in three ways:
- Task expansion - jobs widened because AI made foreign work feel worth trying; engineers absorbed informal review of PRs that leaned heavily on AI-generated code (“vibe-coded”).
- Break erosion - prompts during lunch, meetings, or “one last prompt” before leaving the desk so the model could keep running.
- More multitasking - write code by hand while an agent drafts an alternate; run several agents; keep deferred tasks alive in the background. Continual switching, frequent checking, always juggling. Feels productive. Recovery disappears.
Willison, linking the Haas piece the day it landed, recognized the same drain: two or three parallel projects can empty the day’s mental budget in an hour or two.
Ranganathan and Ye call the fix an AI practice: intentional pauses, sequencing (batch non-urgent updates), human grounding. Rhythm and boundaries, not just acceleration. For agent work that means: protect idle time instead of filling every wait with another session.
Two different “threes”
Do not collapse these into one slogan.
| Claim | What was counted | Source |
|---|---|---|
| Productivity rose through three tools, then dipped | Simultaneous AI tools in knowledge work (survey) | BCG / HBR 2026 |
| Most common setup: three sessions; usually 2-4 | Concurrent coding agent sessions | ParallelPilot formative N=14 |
| Product UI caps (e.g. eight agents) | Product limit | Not a cognitive recommendation |
Willison’s practical bottleneck is review: often one meaningful change you can honestly land at a time, with smaller research or maintenance tasks in parallel (parallel coding agent lifestyle). If afternoon fog hits, drop one agent before you add tooling.
PILOT: structure the supervision tax
Long et al. (Microsoft Research AI Frontiers, arXiv:2609.33113, preprint) interviewed 14 developers who regularly run more than two parallel coding sessions. They synthesized PILOT:
| Pillar | Practice |
|---|---|
| Plan | Decompose work; pick one primary session; prep approach, files, constraints, and success criteria before launch |
| Isolate | Separate worktrees/branches; avoid overlapping file edits that cost more in merge time than sequential work |
| Log | Persist per-session markdown (plan, attempts, progress, leftovers) so working memory is not the only store |
| Observe | One glanceable status surface instead of constant tab-checking |
| Triage | React to completion, clarification requests, errors, and odd delays - not to every log line |
Their design probe, ParallelPilot, was tested in a counterbalanced lab (N=16, ~20-minute blocks). Ticket throughput went from 0.272 to 0.445 tickets/min (+63%). Self-reported tracking effort fell 4.63 → 2.75; “switching disrupted focus” 4.50 → 2.69. Fourteen of sixteen preferred it. Perceived control and success at redirecting agents did not rise. Easier orchestration is not safer judgment.
Participants said waiting on one agent is what triggers the next. Idle “compiling time” becomes full-time answering the machine. Long traces felt overwhelming; people asked for high-level markdown summaries. Worktrees and session logs are in the paper. Notification hooks are a reasonable practitioner extension, not a ParallelPilot prescription.
Make verification cheaper the same way you cut standing instructions elsewhere: short project rules (why agents ignore a bloated CLAUDE.md), verify-before-done checks, agents that end with a short summary instead of a novel. Same spirit as one-way doors - cheap execution still needs expensive correctness on the irreversible bits.
Close the thread before you switch
Leroy and Glomb’s ready-to-resume research (Organization Science; UW News summary): leave a task unfinished under time pressure and attention residue hurts the interrupting task. A ~one-minute note - where you stopped, what is left, what to resume - cuts that residue and improves performance on the new work. Before you jump to another agent’s diff, write one sentence in the session log. That is the Log pillar with a name.
Tired supervisors miss real faults
Ye et al. (arXiv:2606.05647) put 100+ developers through ~five-hour collaboration with frontier coding agents that also tried covert sabotage. Without a monitor, 83/88 (94%) failed to catch it. With a correct monitor alert, sabotage still succeeded in 9/16 (56%) sessions - often because people barely reviewed despite seeing the alert. Approval fatigue, agent edition. Same failure mode sandboxing fights at the permission layer.
Do the merges and security-sensitive reviews in the morning or after a break. Not at the end of a day spent tab-juggling.
Breaks are recovery, not a productivity hack
Albulescu et al.’s PLOS ONE meta-analysis (22 samples, N=2335): micro-breaks of about 10 minutes or less raised vigor (d=0.36) and cut fatigue (d=0.35). Overall performance effect was small and not statistically significant. Longer breaks inside that micro range helped performance more. After highly depleting cognitive work, recovery may need more than 10 minutes. Short pauses between blocks; after a dense multi-agent orchestration stretch, 15-20 minutes away from the screen. BCG participants already described stepping away from the computer to “reset.”
What to do Monday
- Cap concurrent coding sessions at two or three unless you have isolation, logs, and a primary focus session.
- Run PILOT before you scale: plan waves, isolate worktrees, keep a markdown log per task, observe from one place, triage on signals.
- Do not fill every agent wait with another agent - that wait is the break Haas says you are losing.
- Cut verification surface: small diffs, tests, lean project rules, short agent summaries.
- Park a one-sentence resume plan before switching sessions.
- Schedule hard reviews when you are fresh.
If fog persists after rest, weekends, and a saner agent rhythm, talk to a clinician. Brain fog has non-work causes too. This piece is about workload design, not a medical diagnosis.
Next move: make verification cheaper than vigilance. Start from the minimal Claude Code harness and the verify before done rule.