Operators orchestrate Claude, Codex, Hermes on Raft
Machina outlines a multi-agent workflow combining Claude Code, Codex, and Hermes as persistent teammates in a shared workspace called Raft. Running on a local daemon, these specialized agents collaborate in Slack-like channels with compounding memory to build tools, write code, and review each other's work.
Single-assistant chats are a productivity trap; real scale happens when you turn AI models into specialized, self-policing coworkers operating in a unified, local-first workspace.
* Separation of Concerns: Routing jobs to the engine best suited for them (Claude for writing, Codex for building, Hermes for background jobs) avoids model bias and maximizes efficiency.
* The Reviewer Model: Requiring a second agent to review and critique a draft dramatically improves output quality (by up to 90.2% according to Anthropic research) and reduces human oversight.
* Self-Improving Memory Loop: Rejections and feedback are automatically written back to the agents' soul and memory files, ensuring the system naturally adapts and prevents repeat mistakes.
* Privacy and Cost-efficiency: Local execution keeps data secure, while routing routine runs to cheaper API tiers and using existing subscriptions controls costs.
DISCOVERED
4h ago
2026-07-18
PUBLISHED
4h ago
2026-07-18
RELEVANCE
AUTHOR
EXM7777