Agent Harnesses: What They Are & How Hermes, Codex, Claude Code and DeepSeek Harness Fit Together
The software layer that turns a language model into an agent, and where four harnesses sit in the same category.
What is a harness?
An agent harness is the software layer that wraps a language model and turns it from a text predictor into an agent that can act in the real world.
DeepSeek's own framing is Agent = Model + Harness. The model is the soul; the harness lets the agent understand its environment, use tools, and keep working in real-world settings (source).
A harness provides: tool calling (file read/write, shell, web), memory and context management, planning, subagents, sandboxing, session logs, and a UI, terminal or web (source).
Without a harness, an LLM is just a chat box. With one, it can edit code, run commands, browse, and complete multi-step tasks.
Every major agent product is a harness: Claude Code, Codex, Hermes, OpenClaw, and now DeepSeek Harness (source).
Harness vs Codex vs Claude Code
"Harness" is the category. Codex and Claude Code are specific branded products in that category, like "car" vs "Toyota" and "Honda".
| Product | Maker | Models | Interface | Open source | Standalone? |
|---|---|---|---|---|---|
| Hermes | Nous Research | Any LLM (OpenRouter, Anthropic, OpenAI, DeepSeek, local, 20+ providers) | Terminal CLI + Telegram/Discord/WhatsApp gateway | Open source | Yes, fully standalone |
| Claude Code | Anthropic | Claude models (Sonnet/Opus/Haiku) | Terminal TUI | Closed source | Yes |
| Codex | OpenAI | GPT models | Terminal CLI | Closed source | Yes |
| DeepSeek Harness (dsh) | DeepSeek AI | Any model via plugins | Web UI (localhost:3080) | Open source (MIT) | Yes |
Can Hermes connect to Codex or Claude Code?
Yes, two distinct ways.
1. Hermes as orchestrator (delegation). Hermes spawns Codex or Claude Code as subprocesses via its terminal, gives them a task, monitors progress, and relays results. Dedicated skills (codex, claude-code) document the exact commands: codex exec 'task' for one-shots, claude -p 'task' print mode, or tmux for interactive multi-turn sessions. This is how Hermes runs parallel coding agents on separate worktrees (source).
2. Hermes as provider consumer. Hermes can use OpenAI Codex as a model provider (model.provider: openai-codex with OAuth via hermes auth add openai-codex). Anthropic models are also a first-class provider (source).
Does Hermes have its own internal harness?
Yes. Hermes IS a harness. It does not need Codex or Claude Code to function.
- Hermes ships its own agent loop, 40+ tools (web, terminal, file, browser, vision, TTS, cron, delegation), persistent memory, skills system, MCP client, and a multi-platform gateway (source).
- It runs standalone with any LLM provider, including DeepSeek. This very session runs on DeepSeek V4 Flash via ollama-cloud.
- Codex and Claude Code are optional delegation targets for specialized coding work, not dependencies.
Do we need to configure?
Yes, if you want Hermes to delegate to Codex or Claude Code. Current status on this Mac (checked 14 Aug 2026): neither CLI is installed.
Required setup, with user permission (install gate):
- Codex:
npm install -g @openai/codex, thencodex login(orOPENAI_API_KEY). Must run inside a git repo (source). - Claude Code:
npm install -g @anthropic-ai/claude-code, thenclaude auth login(browser OAuth for Pro/Max, orANTHROPIC_API_KEY) (source). - No config needed for Hermes itself. It calls the CLIs via terminal. Optionally
hermes auth add openai-codexto use Codex as a model provider.
DeepSeek Harness (dsh), the new one
DeepSeek Harness is DeepSeek AI's open-source agent harness, released in developer preview (Aug 2026). MIT license, ~39.8k GitHub stars within days (source, source).
- "Everything is a plugin" architecture, powered by the Cordis kernel. Every capability, models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI, is a plugin that can be swapped or recomposed via configuration, without touching source code (source).
- Four runtime modes: Standard (full toolset: file editing, shell, web search, skills, planning, subagents), Code mode (tools exposed via a Code Mode SDK so the model writes TypeScript programs to orchestrate multi-step operations), Minimal (two tools only: persistent bash + str_replace_editor, for benchmarking models), and Creator mode (inspect runtime, test plugins, build custom presets) (source).
- Every run is traceable: append-only session logs record everything the model sees (system prompts, reasoning, tool calls and results, subagent scheduling). A trajectory view lets you inspect, resume, fork, search, and replay (source).
- Run it:
npx @deepseek-ai/dsh webopens a Web UI athttp://127.0.0.1:3080. Node.js required (source). - Status: developer preview, iterating rapidly, compatibility-breaking changes expected. Not production-stable (source).
There is no native Hermes↔dsh integration (no provider entry, no MCP server, no documented bridge) as of 14 Aug 2026. dsh is a harness, not a model API, so it can't be added as a model provider. Like Codex or Claude Code, it could be run as an external CLI process from Hermes' terminal (npx @deepseek-ai/dsh ...), with Hermes orchestrating it. The DeepSeek models themselves are already first-class in Hermes via the DEEPSEEK_API_KEY provider, so you don't need dsh to use DeepSeek models with Hermes (source).
Recommendation: watch it, don't build on it yet. Dev preview plus breaking changes. Re-evaluate when it hits stable.
Bottom line
- A harness is the body around the model brain. Codex, Claude Code, Hermes and dsh are all harnesses.
- Hermes is a full harness itself AND an orchestrator that can drive Codex and Claude Code as subagents.
- To enable that: install and auth the two CLIs (needs the buddy's go-ahead).
- DeepSeek Harness has an impressive architecture (everything-is-a-plugin), but it is a dev preview with no native Hermes integration yet. DeepSeek models already work in Hermes today.