Can My Agent Uselookup · compare · cite

Aug 28, 2026 · Catalog foundation is up

Product feature · runtime

Subagent permissionsCommon product term

Control child-agent tools, inherited permissions, and approval behavior.

Terminology basis: Common product term. Anthropic — Create custom subagents

Subagent permissions: 4 supported, 2 partial, 0 unsupported, 25 unreviewed across 31 cataloged products.

Markdown · JSON

Explore this familyMore in Subagents9 capabilities

Current evidence by product

Can my agent use Subagent permissions?

Read across for the answer. 6 of 31 current product columns have reviewed evidence; unreviewed does not mean unsupported.

Known issues (0)Report an issue
  • Supported4
  • Partial2
  • Unsupported0
  • Unknown25
  • Not applicable0

Web

9 products

Desktop

13 products

CLI

9 products

Unknown means no public evidence has been reviewed for that product and capability. It does not mean unsupported.

How statuses are assigned

Definition and scope

What this capability means

This row asks which tools and permissions a child receives and how its actions obtain approval. A child may inherit parent permissions, receive a reduced tool set, prompt the operator, ask the parent to broker an action, or be unable to request more access.

Evidence should cover one-time and persistent approvals, resource and command scope, MCP and connector credentials, filesystem and network access, nested descendants, denial propagation, audit identity, and whether the operator can review a combined action plan before fan-out.

Traceable compatibility

Assertion ledger

Documentation evidence only. No runtime conformance test is implied.

ChatGPTweb · current
Partial
Target
2026-08-28 ChatGPT Work web documentation observation · hosted-observation
Environment
hosted-default
Observed
2026-08-28
  • policychildren use the tools available to the parent chat and website or connector permissions remain tool-specific; no local sandbox or approval-mode control is exposed
Evidence
ChatGPTdesktop · current
Supported
Target
current ChatGPT desktop documentation · dated-documentation
Environment
local-default
Observed
2026-08-28
  • policysubagents inherit the permission mode chosen for the parent turn; choose it before delegation
Evidence
Codex CLIcli · current
Supported
Target
current Codex CLI documentation · dated-documentation
Environment
local-default
Observed
2026-08-28
  • policyinteractive approval prompts identify the child thread; actions needing a new approval fail in non-interactive execution and return the error to the parent
Evidence
Gemini CLIcli · current
Supported
Target
current Gemini CLI documentation · dated-documentation
Environment
local-default
Observed
2026-08-28
  • policytool lists, inline or inherited MCP servers, and named policy rules can be scoped to a child; recursion protection always removes other agents
Evidence
Claude CLIcli · current
Supported
Target
current Claude Code subagent documentation · dated-documentation
Environment
local-default
Observed
2026-08-28
  • policyparent `bypassPermissions`, `acceptEdits`, or auto mode can take precedence over a child definition; background subagents auto-deny actions that would require a new prompt
Evidence
Copilot Chatdesktop · current
Partial
Target
current VS Code subagent documentation · dated-documentation
Environment
local-default
Observed
2026-08-28
  • policychild identity and tool allowlists are configurable, but separate child approval ownership, grant consumption, and escalation semantics are not fully documented
Evidence
  1. 1. Evidence checked 2026-08-28: Claude Code documents child-specific tool allowlists and denylists plus `default`, `acceptEdits`, `auto`, `dontAsk`, `bypassPermissions`, and `plan` permission modes, with explicit rules for parent-mode precedence.
  2. 2. Evidence checked 2026-08-28: ChatGPT Work children use the parent chat's available tools; local Codex children inherit the selected permission or sandbox policy, surface interactive approvals by source thread, and fail actions that need fresh approval in non-interactive flows.
  3. 3. Evidence checked 2026-08-28: Gemini CLI subagents can receive explicitly scoped tools, isolated MCP servers, and subagent-specific policy files; other agents remain unavailable to descendants.
  4. 4. Evidence checked 2026-08-28: VS Code coordinator agents can allowlist callable children, and each custom child can restrict built-in, extension, and MCP tools; the reviewed docs do not define a separate child approval ledger.