How Many AI Agents Can You Actually Manage?
Pavel Jiranek · 18 November 2026
Read the originalPavel Jiranek, who runs up to nineteen AI coding agents during a release, argues that the work awaiting a person's review sets how many agents that person can manage.
Pavel Jiranek, who writes the newsletter The Hybrid Leader, described how he supervises AI coding agents in his own software work. A quiet day involves two agents, and release preparation can involve up to nineteen that fix bugs, test, write documentation and review changes. He finds that his limit shows when he returns to a task and its context no longer comes easily, and he reviews early research on the load of supervising agents.
- Wu and colleagues surveyed 203 users of AI coding agents, and 83% reported a lighter workload. Of the 203, 101 also rated at least five of ten supervision tasks as very or extremely mentally demanding.
- Usage data from Jellyfish Research for Claude Code, an AI coding tool, found that about 84% of users peaked at one or two sessions running at once over 30 days.
- Jiranek separates four ways of working: guiding one agent, running separate tasks in parallel, comparing several agents' answers to one question, and passing one task through staged roles.
- He advises counting unfinished outcomes, the items awaiting judgement and the time needed to return to a task. He notes that the studies are early and set no safe number of agents.
Teams whose developers run several AI agents can plan review time and authority as deliberately as the coding work. Jiranek's reason is that one developer's extra pull requests become a colleague's checking. Only the open first half of the article was read; the practical guide that follows it requires a free subscription.
Before starting another AI coding agent, count the agent tasks that are finished or paused and still need the developer's review or decision.
When a team plans its work, set aside review time for pull requests produced with AI coding agents and name who has authority to approve them.
Derived by Working Surface from the article; more in the Playbook. Source line: The number of agents I can start is not the number of outcomes I can reliably bring to completion.
22 November 2026: as AI agents make building cheap, the check before production limits the pace, so teams are redesigning who sponsors early builds, what evidence a change must carry, how agent work is batched for testing and how much one person reviews.