← All Skills
🤝
Agent Collab Governed Multi-Agent Skills
Seven provider-neutral skills for task splitting, context budgets, plan-act-reflect loops, output reconciliation, adversarial debate, proposal-only shared memory, and evidence acceptance while preserving final human authority.
Key Features
- DAG task packets and role boundaries
- Policy-driven context and checkpoints
- Failure-preserving reconciliation and acceptance evidence
- Proposal-only memory and human shipping authority
Installation
In Claude Code, add the WenyuChiou/agent-collab-skills marketplace and install agent-collab-workspace@agent-collab-skills. The optional deterministic validation layer is installed with python -m pip install agent-collab-harness.
How to Use
For work needing multiple bounded roles, use task-splitter to produce a provider-neutral DAG and acceptance criteria, then apply context-budget, plan-act-reflect, reconciler, and acceptance-gate to manage execution evidence. The skills describe collaboration; the host supplies models, tools, sessions, and sandboxes.
Requirements
The skill bundle targets Claude Code; the optional harness requires Python and local strict-JSON policies/checkpoints. Human decisions and overrides may use an HMAC key supplied by a trusted host; never expose it to child agents or commit it. .coord/.ai are local scratch by default. Memory changes, exceptions, and shipping need separate human decisions; install alone authorizes no spawn, commit, merge, or push.
Related Skills