Community fork: per-agent isolation for multi-agent OpenClaw deployments

[OctoClaws]

👋 Community Contribution Notice

We maintain a fork with multi-agent isolation enhancements for OpenClaw deployments with multiple agents:

Fork: https://github.com/OctoClaws/MetaClaw

What it adds

When running multiple agents (e.g., research assistant, code reviewer, daily assistant) through a single MetaClaw instance, the default behavior shares one skill library across all agents. This causes cross-contamination — skills evolved from research conversations get injected into casual chat prompts, degrading quality for both.

Our fork adds:

Feature	Description
Per-agent skill directories	Each agent gets skills/{agent_id}/ + shared skills/_shared/
Per-agent mode routing	Agent A can use skills_only while Agent B uses rl
Per-agent skill evolution	SkillEvolver writes to the correct agent's directory
Per-agent LoRA training	Separate checkpoints per agent (no trajectory mixing)

Implementation

• ~145 lines changed across 3 files
• Fully backward compatible (single-agent setups unchanged)
• Uses X-Agent-Id header from OpenClaw's plugin SDK (ctx.agentId)
• PR submitted: feat: per-agent isolation for skills, mode, and evolution #50

Who benefits

Anyone running 2+ agents through MetaClaw who noticed skills from one agent polluting another's prompt injections.

We'd love feedback from the maintainers and community. Happy to iterate on the PR.