Hermes Agent Multi-Agent — Parallel Agents That Work Together

Multi-Agent Architecture

Updated: April 4, 2026·hermes agent multi-agentmulti-agentparallel

Run multiple Hermes agents in parallel — orchestrator + worker patterns, when it makes sense, and when one agent is enough.

Multi-agent setups are where Hermes surpasses ChatGPT, Claude Code, and other single-agent tools — Hermes gets genuinely powerful — spawning separate agents for different responsibilities. Here is how to set them up and common patterns.

Why Multi-Agent

Single agents bottleneck on complex work. Multi-agent lets you parallelize:

Research agent + execution agent
Orchestrator + validator
Different specialists for different domains

One Discord user: "I think I have truly entered the multi terminal stage of my AI journey."

Patterns

Orchestrator + Worker

One agent coordinates, delegates subtasks to workers.

Supervisor + Subagent

Main agent reviews work, spawns subagent for specific tasks.

Specialist Agents

Separate agents for different domains:

Coding agent
Research agent
Scheduling agent

Example from Discord:

"I'm using droid and have kimi as an orchestrator and worker but minimax as validator"

This is kimi (cheap, fast) coordinating work, minimax (better reasoning) validating outputs.

Spawning Subagents

In conversation: "Research X in a subagent, use the research agent" or "Spawn a research agent to look into this"

Each subagent has:

Own conversation thread
Own terminal
Own Python RPC scripts
Zero-context-cost pipelines via execute_code

Multi-Agent Commands

List active agents:

hermes agents list

Switch between agents:

hermes agents use [agent-name]

Use Cases

Parallel Research

Main agent receives task
Spawns 3 research agents for different topics
Each works in parallel
Main agent synthesizes results

Code + Validate

Coding agent writes code
Validation agent reviews for bugs
Loop until validated

Team Simulation

Multiple agents with different personas collaborate on a task.

Configuration

Edit ~/.hermes/config.yaml:

multiagent:
  enabled: true
  max_agents: 5
  default_role: worker  # orchestrator or worker

Skills system Multi-agent feature

FAQ

Does each agent count as separate API calls? Yes. Each spawns its own conversation.

Can agents share memory? Currently, each agent has isolated memory. Shared skills are separate.

Set up multi-agent workflows

Skills guide | Community projects

flyhermes.ai

Frequently Asked Questions

When should I use multi-agent setup instead of a single Hermes instance?

Use multi-agent when you have parallel workflows that can run independently — such as a research agent and an execution agent working simultaneously, or separate specialists for coding, research, and scheduling. Single agents bottleneck on complex work; multi-agent lets you parallelize without context bloat.

Does spawning multiple agents multiply my API costs?

Yes. Each subagent spawns its own conversation thread and makes its own API calls. This is why community members recommend using a cheap fast model (like Kimi K2.5) for the orchestrator and a more capable model (like Claude) as a validator — quality control without running expensive models on every subagent.

How do I actually spawn a subagent from a conversation?

Just ask: 'Research X in a subagent' or 'Spawn a research agent to look into this.' Each subagent gets its own conversation thread, own terminal, and zero-context-cost pipelines via execute_code. Use `hermes agents list` to see active agents and `hermes agents use [name]` to switch between them.

Can multiple agents share the same memory or skills?

By default, each Hermes agent has isolated memory. Skills are shared if they're in the shared ~/.hermes/skills/ directory. For true shared episodic memory across agents, the PLUR community plugin enables engram sharing — corrections made to one agent propagate to others on the same project.

What is a practical multi-agent pattern for a content team?

One Discord user described running a social media manager agent writing posts, a visual agent generating branded images, and an ads creator agent consuming both outputs — all coordinated by a Kimi-powered orchestrator with a MiniMax validator reviewing quality before anything ships.

Ready to Run Your Own AI Agent?

Self-host Hermes in 60 seconds. No credit card, no cloud lock-in.

Deploy Hermes Free →