Idle-agent anxiety has a formula: work time over instruction time
Mar 27 essay by Eduardo Aguilar Pelaez names the low-grade dread of an agent waiting on you, models it as a queue with the human as the server, and argues the lever is longer autonomous work intervals plus an orchestrator, not more agents.
Source: The Agent Tending Problem: Why Your Coding Agents Need an Orchestrator, Not More Agents
Eduardo Aguilar Pelaez, CTO at Legal Engine, starts from a feeling most people running parallel coding agents will recognise: the background anxiety of knowing one of your agents is sitting idle right now because you haven't gotten to it. He calls it idle-agent anxiety, and says it is worse than ordinary multitasking stress because the cost of your inattention is visible and growing.
His model is a queue with you as the server. The number of agents you can usefully run is roughly N = T_work / T_instruct: how long an agent works on its own before needing you, divided by how long it takes you to read its output, decide, and give the next instruction. Ten minutes of autonomous work against two minutes of your attention supports about five agents. He caps it in practice at around five to seven, where keeping a mental model of each workstream falls apart.
Two conclusions follow. The bigger lever is stretching T_work (better context, clearer specs, permission to make safe decisions without asking), because the time you spend instructing has a floor and the time an agent can run doesn't. And the answer to too many tabs is not more agents but a supervising layer, his analogy being CNC machine tending: durable task state, a supervisor that restarts stalled agents, policy on what each agent may touch, and routing your attention to whichever agent is blocked. He's candid that his own runtime for this (Elan, built on Erlang's BEAM) exists only in part; the attention-routing pieces are still planned.
Why it matters for code review
This puts a number on the emotional half of running agents around the clock. The urge to keep every agent busy is real, but the formula says the thing worth optimising is total useful work per unit of human attention, which often means fewer agents running longer, not more agents nudged more often. Merge conflicts and incompatible changes, he notes, grow faster than fleet size. Review is that attention budget too: every verified-by-machine decision lengthens T_work, and every PR that needs a human read shortens it.