Using Claude Code cross-session messaging as a multi-agent team: Is Anthropic inching toward a Hermes-style runtime?
retained text
Ever since Claude Code introduced cross-session messaging (`SendMessage`/`ListAgents`), I’ve stopped running multi-step development sequentially in a single terminal. Instead, I spin up multiple terminal tabs using `--name` and `--dangerously-skip-permissions`, and ask the orchestrator to handshake with the named session and assign each a dedicated role (by asking them to read their respective runbooks) : * `Orchestrator-Agent` * `Researcher-Agent` * `Content-Writer-Agent` * `Developer-Agent` * `Deploy-Agent` **Why this workflow works well:** 1. **Parallel Execution:** They communicate and hand off findings directly over local sockets without me manually copy-pasting context between terminals. I only talk to the Orchestrator. 2. **Context Window Hygiene:** Instead of letting one long-running session degrade as the context window blows up, each session maintains a clean, domain-specific context. **The Hermes connection:** This setup feels like it's approaching what **Hermes Agent** does. Obviously Hermes provides an always-on daemon, long-term memory, scheduling, and autonomous background execution. But Claude Code already has the primitives: local IPC, agent discovery, bash toolkits, and permission boundaries.
document fetched 2026-09-30 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wtdfg1/using_claude_code_crosssession_messaging_as_a