Agents Managing Agents

Verdict: thin. This is term 22 with a longer leash and worse observability. A worker that can dispatch workers does not change the shape; it changes who chose the depth, and the answer becomes nobody.

The only genuinely new content is two mechanisms, so the unit is those two.

the tree: 12 agents, 42500 tokens if every one runs
nobody wrote "12 agents" anywhere. The planner delegated, and so did the leads.

A shared pool, not a ceiling each. A per-agent limit of 5,000 tokens is satisfied by every one of those twelve agents while the tree spends 42,500. A ceiling each is not a budget. One pool, decremented before the call:

one shared pool of 20000:
  completed 6 of 12 agents
  stopped at w:verify-1: requested 4000 tokens with 0 left; refused before the call

A kill switch that nothing can clear. Anything may trip it; it stops the tree before the next unit of work rather than after the current one; and the first reason wins, because a later kill overwriting it destroys the only account of what went wrong.

a kill switch tripped mid-run:
  completed 4: planner, lead:extract, w:extract-1, w:extract-2
  stopped at w:extract-3: stopped by output-watcher: the second extractor
    returned a field it had marked unreadable

A depth limit is the third thing and it is really the first one restated. An unbounded tree with a shared pool does terminate — when the pool empties — which is a termination condition made of money. A number is easier to reason about at 3am.

Every spend is attributed in a ledger, because the question after a runaway is always which one.

Sharp edge

An agent here is a node with a token cost. Nothing in this run is a model, so nothing here says whether a supervising agent would delegate well — only what the supervision mechanisms do once it has decided.

Break it first

mutation                                  tests                 demo
the pool is per agent, not shared         1 failed | 11 passed  exit 1
a later kill overwrites the first reason  1 failed | 11 passed  exit 0
the kill is checked after the spend       1 failed | 11 passed  exit 0
depth is recorded but not enforced        2 failed | 10 passed  exit 1

"The kill is checked after the spend" leaves the demo green because the tree still stops — one unit of work later, having paid for it. That is the whole difference between a kill switch and a notification.