Comparative framework across Claude Code, Hermes, and OpenClaw for multi-agent orchestration. Ephemeral sub-agents vs persistent agents, when to use each, and which harness is best for which type of work. Validates the Carlos fleet architecture from an independent external perspective.

Watch on YouTube

The Three Patterns

Sequential (Chatbot Mode)

One agent, one session. You prompt, you get the answer. Limited to context window. What most people do.

Fan-Out (Ephemeral Sub-Agents)

Main agent spawns parallel sub-agents. Each gets clean context, does one task, returns a summary, disappears. Main agent synthesizes.

Examples: research agent, design agent, title agent run simultaneously, return results, main agent combines output.

Persistent Multi-Agent Teams

Named agents with their own memory, workspace, soul, and identity. Feel like teammates. Run continuously, remember past interactions, build on prior context.

How Each Harness Handles It

Claude Code

• Sub-agents defined in claude.ai/agents as instruction files
• CLAUDE.md instructs orchestrator to delegate to sub-agents
• Sub-agents get clean context, do work, return output, disappear
• Key benefit: keeps orchestrator window clean and reduces cost
• Limitation: Anthropic model ecosystem primarily (workarounds exist but advanced)
• Best for: one-off tasks, code shipping, file processing at scale

Hermes Agent

Two delegation mechanisms:

Delegate task (ephemeral): equivalent to Claude Code sub-agents. Spawn, work, return summary, vanish. No memory. Use for: parallel one-off tasks, fan-out to 5 things simultaneously.

Profiles (persistent): each profile has own soul, memory, model, gateway, auth. Equivalent to OpenClaw workspaces. Use for: recurring workflows, role-specific memory, different models per role.

Hermes advantage: best out-of-box memory. Remembers everything from days ago without configuration.

Hermes caveat: self-improving skills can proliferate into dozens of auto-created skills you did not intend. Monitor skill count.

Trigger a profile: hermes chat -p [profile-name] -prompt "[task]"

OpenClaw

• Workspaces are more fully scoped than Hermes profiles: can include examples, scripts, tools
• Each workspace has own soul, memory, and everything that agent needs
• Workspaces not isolated by default (can be configured but clunky)
• Supports sub-agent depth: agent spawns another agent which spawns 10 sub-agents for variations
• Feels most like a real team with identity, names, collaboration
• Memory requires setup work (Obsidian, Pinecone, or other database)

Nemanja switched from OpenClaw to Hermes specifically for the memory system. OpenClaw requires building and wiring memory. Hermes has it built in.

Decision Framework

Task Type	Recommended
Pure code shipping, one-off tasks	Claude Code
Analyze 2,000 transcript files	Claude Code
Daily automations, content pipelines	Hermes or OpenClaw
Named team with identity and memory	OpenClaw (or Hermes + OpenClaw)
Persistent memory, self-improving workflows	Hermes
Orchestrating Claude + Codex workers	Hermes as orchestrator

Sub-Agents vs Persistent Agents

Sub-agents (ephemeral):

• Spawn for one task, return summary, vanish
• No memory between runs
• Like hiring a freelancer: completes the job, forgets you
• Best for: parallel processing, fan-out, one-off research

Persistent agents:

• Long-lived, named, have own memory
• Build context over time
• When they need parallel work: spawn sub-agents underneath

When persistent agents need to do parallel work, use sub-agents under them. The two patterns are complementary, not competing.

Hermes + OpenClaw Together

Hermes can orchestrate an OpenClaw agent team. OpenClaw agents execute as persistent specialists. Hermes provides memory layer, loop management, messaging integration. Best-of-both-worlds for operators who want full team identity AND good memory.

Carlos Fleet Validation

This video validates the Carlos fleet architecture from an external perspective:

• Carlos (conductor) = Hermes orchestrator role
• Specialist agents = Hermes profiles or OpenClaw workspaces
• Agent tool dispatches = Hermes delegate task
• The fleet already correctly separates ephemeral dispatch (one-off) from persistent specialists (recurring)