[BUG] Agent tool spawns 7 named teammates for a single routine task; nested spawning burns 77% of 5h limit in <1 hour

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Code v2.1.207, Opus 4.8, Max 5x plan, macOS.
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS enabled.

A multi-step task issued 7 Agent tool calls. Six were plain subagents with
explicit cheap models (Explore/Haiku, general-purpose/Haiku, /Sonnet) and
behaved fine.

The 7th call passed a name parameter ("gate-impl"), which makes it a
teammate rather than a subagent. Two things then went wrong:

1. Its agent type (flutter-app-developer) had no tools: restriction in its

definition, so it silently inherited spawn-capable tools (Agent/Task/
SendMessage) — and recursively spawned agents of its own.

1. It had no model: set, so it inherited Opus for every nested agent.

There is no recursion-depth cap, no fan-out cap, and no warning before an
agent tree consumes a large share of the session limit.

Session log evidence:
grep -c '"name":"Agent"' -> 7
grep -c 'TaskCreate|TaskUpdate|SendMessage' -> 57
wc -l -> 1577
du -h -> 7.0M

Result: ~77% of the 5-hour limit and ~8% of the weekly limit consumed in under
one hour. Second occurrence in two weeks. Quota is not recoverable.

What Should Happen?

1. Hard recursion-depth cap on agent spawning (default 1). A spawned agent

should not be able to spawn further agents unless explicitly opted in.

1. Omitting tools: in an agent definition should NOT silently grant

spawn-capable tools (Agent/Task/SendMessage). Spawn tools should be
opt-in, not inherited by default.

1. A guardrail before an agent tree consumes a large fraction of the session

or weekly limit — warn, or require confirmation, rather than silently
burning through it.

1. Ideally: a fan-out cap, and a visible running cost indicator while an

agent tree is active.

Error Messages/Logs

Steps to Reproduce

1. Enable agent teams:

~/.claude/settings.json
{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

1. Define an agent WITHOUT a tools: restriction and WITHOUT a model: field.

.claude/agents/impl-agent.md
---
name: impl-agent
description: Implements application code.
---
You implement code. Report files changed and test results.

(No tools: -> inherits all tools, including Agent/Task/SendMessage.
No model: -> inherits the main session model, i.e. Opus.)

1. Start a session on Opus 4.8 and give it a multi-step task large enough that

it decides to delegate — e.g. "investigate these two repos, plan a fix,
file the tickets, then implement the fix in a worktree."

1. Let the lead delegate. It will issue several plain subagent calls (fine),

and at least one Agent call that passes a name parameter — which makes
that one a teammate rather than a subagent.

1. Observe: the named teammate, having inherited spawn-capable tools, spawns

further agents of its own. There is no depth cap and no fan-out cap.

1. Inspect the session log:

SESSION=$(ls -t ~/.claude/projects//.jsonl | head -1)
grep -c '"name":"Agent"' "$SESSION" # -> 7
grep -c 'TaskCreate\|TaskUpdate\|SendMessage' "$SESSION" # -> 57
wc -l "$SESSION" # -> 1577
du -h "$SESSION" # -> 7.0M

The Agent call with a name field is the one that cascades:
{"subagent_type":"flutter-app-developer","name":"gate-impl","prompt":"..."}
The other six carry an explicit cheap model (haiku/sonnet) and no name,
and they behave correctly.

1. Check /cost or the limit indicator: ~77% of the 5-hour limit and ~8% of the

weekly limit gone in under an hour.

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.207

Platform

Other

Operating System

macOS

Terminal/Shell

Other

Additional Information

Agent definitions in use (.claude/agents/):

• flutter-app-developer : no tools:, no model: <- the one that cascaded
• (others: Explore, Plan, general-purpose — built-in, behaved correctly)

The six well-behaved Agent calls each passed an explicit cheap model
(haiku / sonnet) and NO name parameter. The single cascading call passed
name: "gate-impl" and no model.

Two independent contributing factors, either of which alone would have
limited the damage:
(a) name param promotes the call to a teammate, which retains spawn tools
(b) omitting tools: in the agent definition grants Agent/Task/SendMessage

Not a one-off: a comparable runaway burn occurred the previous week on an
earlier 2.1.x, though I did not capture logs at the time.

Full session log available (7.0MB); can provide the 7 Agent invocations
as a redacted excerpt on request.