Task(AgentName) Disable

Disabling specific agents through Task(AgentName) syntax blocks selected subagent types at availability time, filtering them from Claude's agent option set without throwing errors.

How it works

The --disallowedTools CLI flag and settings.json permissions support both broad and granular blocking:

• Task alone blocks all subagents entirely.
• Task(Explore) blocks only Explore-type agents; the Task tool remains available and other agent types (Plan, Bash, general-purpose) continue available.

Disabling works at the availability level: blocked agents do not appear in the set of agent types Claude can select from, so they never become invocation candidates.

What the test found

Three scenarios were verified on 2026-01-07:

1. Blocking all subagents (Task) removed the tool entirely. Pass.
2. Blocking a single agent type (Task(Explore)) filtered only that agent, leaving Task and other types functional. Pass.
3. When requesting a disabled agent, Claude did not error. Instead it worked around the constraint by selecting alternative agents or direct tools, demonstrating graceful degradation.

The syntax scales: Task(Plan) would block Plan agents without affecting others, though Plan was not explicitly tested.

Why it matters

This gives three concrete capabilities:

• Cost control: Disable expensive agents (Opus-based types) to constrain spending.
• Security: Prevent specific agent capabilities where policy forbids them.
• Testing: Isolate agent behaviors by removing interference from other types.

The graceful fallback means requests never fail catastrophically; Claude adapts rather than stopping.

Mechanics

The availability filter operates before agent selection. Claude's agent-type decision tree sees a reduced set of options based on the disallowedTools configuration. This differs from a hard runtime block and explains why no errors surface.