Agent Collaboration Framework
PassAudited by ClawScan on May 10, 2026.
Overview
This is a transparent instruction-only framework for coordinating multiple agents, but it broadens inter-agent communication and relies on persistent sessions and handoff files that users should supervise.
Install this only if you intentionally want multiple OpenClaw agents to collaborate. Review the openclaw.json and AGENTS.md changes before applying them, keep allowAgents lists narrow, avoid secrets in HANDOFF files, and monitor or close persistent sessions when collaboration is finished.
Findings (3)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Information sent to one agent may be available for cross-agent collaboration depending on the user’s OpenClaw configuration.
The setup asks the user to make sessions broadly visible so agents can collaborate. This fits the skill purpose, but it means context may be shared across agents more widely than a single-session workflow.
"tools": { "sessions": { "visibility": "all" } }Enable broad session visibility only for agents you trust, and restrict collaboration to the specific agents needed for your workflow.
Sensitive details placed in HANDOFF files may be read by the receiving agent, and deletion after reading can reduce auditability.
The HANDOFF workflow stores instructions in local files that another agent later reads and deletes. This is a disclosed fallback mechanism, but users should treat those files as task-bearing context.
写入 ~/.openclaw/workspace-content/HANDOFF-FROM-MAIN.md ... *接收方讀取後刪除此文件*
Keep HANDOFF files limited to intended task details, avoid secrets, and ensure only trusted users or agents can write to the OpenClaw workspace directories.
A poorly scoped request could lead to extra agent activity or duplicated work across departments.
The framework allows department agents to delegate within their own scope without routing every task through Main. This is purpose-aligned, but a mistaken task could propagate to multiple subagents.
**部門主管在自己範圍內自主調度**，不需要經過 Main
Review the allowAgents lists, keep department scopes narrow, and ask agents to confirm before delegating high-impact tasks.