openclaw - 💡(How to fix) Fix Subagents spawned via sessions_spawn cannot use browser, web_search, web_fetch despite alsoAllow config [1 participants]

Problem

Subagents spawned via sessions_spawn cannot use browser, web_search, or web_fetch tools, even when explicitly allowed in configuration. This fundamentally limits multi-agent architectures where a main orchestrator needs to delegate web-based tasks to specialized subagents.

This is a renewed report to give visibility to a long-standing limitation originally tracked in #41577 and #45873, which remain unresolved as of v2026.4.5.

Configuration

{
  "tools": {
    "profile": "full",
    "subagents": {
      "tools": {
        "alsoAllow": ["browser", "web_search", "web_fetch"]
      }
    }
  },
  "agents": {
    "list": [
      {
        "id": "orchestrator",
        "name": "Orchestrator",
        "model": {"primary": "anthropic-provider/claude-sonnet-4-5"},
        "subagents": {"allowAgents": ["*"]}
      },
      {
        "id": "web-worker",
        "name": "Web Worker",
        "workspace": "/install/workspace-web",
        "model": {"primary": "openai-provider/gpt-5.4"}
      }
    ]
  }
}

Expected behavior

When the orchestrator spawns a subagent via sessions_spawn targeting the web-worker agent, that subagent should be able to use browser, web_search, and web_fetch tools — as declared in subagents.tools.alsoAllow.

Actual behavior

The spawned subagent does not have access to browser, web_search, or web_fetch. Only the main (parent) agent can use these tools directly. The subagent either ignores the tool call or reports the tool as unavailable.

Reproduction steps

1. Configure a multi-agent setup with at least two agents (one orchestrator, one worker)
2. Set subagents.allowAgents: ["*"] on the orchestrator
3. Set tools.subagents.tools.alsoAllow: ["browser", "web_search", "web_fetch"] globally
4. Start the gateway and interact via any channel (Discord, Telegram, etc.)
5. Ask the orchestrator to delegate a web task (e.g., "search for X on the web") to the worker subagent via sessions_spawn
6. Observe that the subagent cannot use browser/web tools

Impact

This limitation forces all web-based automation into a single agent, creating significant bottlenecks:

• Web scraping pipelines that should be parallelized across specialized subagents must funnel through the main agent, overloading its context window
• Content research workflows where one subagent should gather data (web) while another processes it (text) cannot be properly separated
• Monitoring workflows where multiple subagents need to check different web sources concurrently are impossible — everything serializes through the parent

The only workaround is routing ALL web tasks through the main agent, which defeats the purpose of multi-agent orchestration and creates context window pressure on long-running sessions.

What has been tried

• Setting subagents.tools.alsoAllow at agents.defaults level — no effect on spawned subagents
• Setting subagents.allowAgents: ["*"] on each agent — does not grant tool access
• Various combinations of tools.profile: "full" with explicit tool allowlists — no change

Related issues

• #41577 — original report of subagent tool restrictions (open since early 2026)
• #45873 — additional report focusing on browser tool availability in subagents

Both remain open with no resolution. The v2026.4.2 fix #59555 (subagent spawn pairing) and v2026.4.5 improvements (#59944, #58178) improved subagent stability but did not address tool availability.

Environment

• OpenClaw v2026.4.5 (3e72c03)
• macOS (Apple Silicon)
• Tested with Discord and web chat channels
• Multiple provider configurations (Anthropic, OpenAI)