One research prompt: 394 Claude Code subagents, 341 at once, 84M tokens in ten minutes. Locked out.
So I built agent-usage-guard, a mod that caps agent fan-out across sessions and asks before heavy context burns your window. Free, MIT.
https://t.co/2zGWSSRZ3g
@HoagiemanApps Ha, honestly it's more impatience than patience. I didn't trust my eyes, so I got a robot to check for me. Bottling Bear's feel is the harder job, there's no test you can run for "nice to write in"
Made an Obsidian theme that's easy to read and easy to scan.
I didn't want to trust my eyes, so Tela is contrast-audited on 346 screens: notes, canvas, bases, menus, settings, dark and light, desktop and phone.
One optional setting. Free, MIT.
https://t.co/fj2TdQtE9R
@mariioalexandre Checked that run. Each subagent's first request was about 29k with the task included, so the starts came to about 10M, an eighth of the total.
The rest was the agents working.
One research prompt: 394 Claude Code subagents, 341 at once, 84M tokens in ten minutes. Locked out.
So I built agent-usage-guard, a mod that caps agent fan-out across sessions and asks before heavy context burns your window. Free, MIT.
https://t.co/2zGWSSRZ3g
@swordBluesy The second scene in the video is that test. Claude starts five agents with a limit of four, and the question comes up before the fifth one starts.
Agents running in other Claude Code sessions on the same machine count toward the four too.
@LankC11 Claude Code added one in July, four weeks after that lockout. It caps each session at 20 subagents running at once (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS).
Nothing limits how many a session starts over time, though, and each session only counts its own.
@Joshua_WD Expensive lesson. Sixty days of logs afterwards showed it wasn't a one-off: 12 lockouts. So the cap had to be on by default, not something you remember to set.
@karanjagtiani04 Exactly, so it lives in a hook: the model asks for an agent, the mod decides. A refusal comes back with a reason the model can act on, so it waits or batches instead of retrying.