| name | hermes-subagent |
| description | Delegate complex, autonomous, or tool-heavy tasks to a Hermes Agent sub-agent running in the AGNT sandbox. Hermes is Nous Research's self-improving Python agent (47 built-in tools, persistent memory, skills system, sub-agent delegation). Use this skill whenever the user asks Annie to "delegate to Hermes", "have Hermes do X", "run this through a sub-agent", "offload this task", "use a sub-agent for autonomous research", "let Hermes handle this", or any phrasing that hands a substantial, multi-step, or tool-heavy task off to an external agent. Also trigger when the user mentions Hermes Agent, Nous Research's agent, the AIAgent class, run_agent.py, or wants to spawn an autonomous sub-process to do research, multi-tool work, code analysis, code execution, or long-running browsing while keeping Annie's main context clean. Encodes the verified working invocation pattern (sandbox path, OpenRouter default, correct API signature, max_tokens cap, common error fixes) so delegation works on the first try. |
| version | 1.0.0 |
Hermes Sub-Agent Delegation
Delegate substantial work to Hermes Agent (Nous Research) running in an isolated AGNT sandbox. Hermes acts as a heavyweight sub-agent: it has its own toolset, persistent memory, skills system, and can spawn its own sub-agents in turn.
When To Delegate vs. Handle In Chat
| Task shape | Where to handle |
|---|
| Quick lookup, single tool call, conversational reply | Annie (this chat) |
| Multi-step research with synthesis | Hermes |
| Long-running autonomous browsing | Hermes |
| Heavy code execution + iteration in a sandbox | Hermes |
| Cross-session continuity (resume work tomorrow) | Hermes (session_id) |
| Anything that would burn 20+ Annie turns | Hermes |
| Image/video generation, AGNT-specific tools | Annie |
If unsure, ask the user one quick clarifier rather than guessing.
Verified Sandbox Layout
Hermes lives in an isolated venv at:
C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\
├── .env # OPENROUTER_API_KEY=...
└── .venv\
├── Scripts\
│ ├── python.exe # use absolute path, never activate
│ ├── hermes.exe
│ ├── hermes-agent.exe
│ └── hermes-acp.exe
└── Lib\site-packages\
├── run_agent.py # AIAgent lives here
└── hermes_agent\ # full package
Never modify the global Python or npm globals. Always call the sandbox python.exe by absolute path. The venv is intentionally isolated from system Python.
Preflight (run once per session before first delegation)
If anything below is missing, fix it before invoking Hermes:
:: 1. sandbox exists?
dir "C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.venv\Scripts\python.exe"
:: 2. hermes-agent installed?
"C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.venv\Scripts\python.exe" -m pip show hermes-agent
:: 3. .env has OPENROUTER_API_KEY?
type "C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.env"
If hermes-agent is missing, install with:
"C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.venv\Scripts\python.exe" -m pip install "git+https://github.com/NousResearch/hermes-agent.git"
If .env is missing or empty, ask the user for an OpenRouter key (or Anthropic / OpenAI / Nous Portal key) and write it as a single line OPENROUTER_API_KEY=sk-or-v1-... to C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.env. Never paste keys into source code.
The Canonical Invocation (this is the right one — others fail)
Use the bundled helper in scripts/delegate.py:
"C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.venv\Scripts\python.exe" "C:\Users\Studio\AppData\Roaming\AGNT\skills\hermes-subagent\scripts\delegate.py" "your task here"
The script:
- Loads
.env from the sandbox.
- Constructs
AIAgent(quiet_mode=True, model='anthropic/claude-sonnet-4.5', provider='openrouter', max_tokens=4000) — these are the verified working defaults.
- Calls
agent.chat(message) and prints the result on stdout.
- Accepts
--model, --max-tokens, --max-iterations, and --toolsets flags.
Why these defaults matter (each one was a real failure mode in testing):
quiet_mode=True belongs in the constructor, not in chat(). The published docs say otherwise. They are wrong for v0.12.0.model and provider must both be set explicitly — without them Hermes' default model resolution returns None silently and your call appears to hang.max_tokens=4000 prevents OpenRouter's pre-flight credit check from rejecting the request as HTTP 402 ("requires more credits"). Hermes defaults to 64,000 reserved output tokens, which is too aggressive for low-balance accounts.
Delegation Patterns
Pattern 1 — Fire-and-forget (most common)
For any single task where you want one final answer:
"C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.venv\Scripts\python.exe" ^
"C:\Users\Studio\AppData\Roaming\AGNT\skills\hermes-subagent\scripts\delegate.py" ^
"Research the top 3 open-source agent frameworks released in 2026 and summarize their differentiators in 200 words."
Run this with _executeAsync: true and an _estimatedMinutes value (typical 1–3 min for ~4k token tasks, 3–10 min for tool-heavy work). The result message comes back automatically when Hermes finishes.
Pattern 2 — Multi-turn with session continuity
When the user wants to continue a Hermes conversation across multiple Annie turns, pass a stable session_id:
... delegate.py --session-id agnt-research-2026-04-30 "follow up: now compare them on cost"
Hermes persists session state in its own SQLite store. Reusing a session id lets Hermes recall prior turns without re-summarizing.
Pattern 3 — Parallel sub-agents
Hermes itself can spawn sub-agents via its delegate toolset. For Annie's purposes, spawn multiple Hermes processes in parallel only when the tasks are genuinely independent. Use _executeAsync: true on each call so they run concurrently. Costs add linearly, so cap at 3–5 parallel runs unless the user explicitly approves more.
Cost & Performance Guidelines
| Task complexity | Typical cost (Sonnet 4.5 via OpenRouter) | Wall time |
|---|
| Simple chat / summary | $0.02 – $0.05 | 30–60s |
| Research + synthesis | $0.05 – $0.15 | 1–3 min |
| Multi-tool autonomous work | $0.15 – $0.50 | 3–10 min |
| Long autonomous loop | $0.50 – $2+ | 10+ min |
For repeated cheap delegation, switch to --model openai/gpt-4o-mini or --model anthropic/claude-3.5-haiku. They handle 80% of sub-agent work at ~10% of Sonnet's cost.
Common Errors And Their Fixes
These are real errors observed during the install/verify session for this skill — handle them inline rather than escalating to the user.
TypeError: AIAgent.chat() got an unexpected keyword argument 'quiet_mode'
Move quiet_mode=True to the AIAgent(...) constructor. The chat() method only accepts message and an optional stream_callback.
agent.chat(...) returns None after a long pause
The model isn't being resolved. Pass both model='anthropic/claude-sonnet-4.5' and provider='openrouter' to the constructor.
HTTP 402: This request requires more credits, or fewer max_tokens. You requested up to 64000 tokens, but can only afford X
Either add credits at https://openrouter.ai/settings/credits, or pass max_tokens=4000 to the constructor (or --max-tokens 4000 to delegate.py). The bundled helper already does this.
ImportError: cannot import name 'AIAgent' from 'run_agent'
You're using the wrong python. Use the absolute sandbox path:
C:\Users\Studio\AppData\Roaming\AGNT\hermes-sandbox\.venv\Scripts\python.exe. Do not use python from PATH.
OPENROUTER_API_KEY env var is empty inside the script
The .env wasn't loaded. The bundled helper calls load_dotenv() against the sandbox .env explicitly. If you're invoking Hermes directly (without the helper), you must replicate that.
Hermes hangs forever, no tokens reported
The model is unreachable or the API key is invalid. Test with a curl to https://openrouter.ai/api/v1/auth/key using the same key. Fix the key in .env.
Reporting Results Back To The User
When you get Hermes' response:
- Quote the substantive output, don't paraphrase. Hermes did the work — show it.
- Include a single status line: model used, tokens in/out, wall time, cost estimate.
- If Hermes used tools, mention which ones briefly (its trace is interesting).
- Offer next steps: refine the answer, dig deeper, or move on.
Format example:
Hermes (sub-agent) returned:
[verbatim output from Hermes]
⏱ Sonnet 4.5 · 3,200 in / 410 out · 47s · ~$0.02
When Hermes Is The Wrong Tool
Don't delegate when:
- The task is a single image/video generation — use Annie's
generate_image or seedance_api directly.
- The task needs an AGNT-specific tool (workflow operations, plugin install, AGNT API calls).
- The task is conversational ("how should I name this?") — Annie is faster.
- The task is < 30 seconds of straightforward work.
In those cases, do it inline and don't burn the Hermes round-trip.
Deeper Reference
For full Hermes API surface (toolsets, run_conversation, callbacks, multi-turn patterns, model selection, MCP integration), see reference.md in this skill folder. Read it when:
- The user wants Hermes to use specific toolsets only.
- You need to inspect Hermes' tool trace, not just the final answer.
- You're integrating Hermes deeper (e.g., scaffolding the AGNT plugin).
- A subtask requires features not covered above.
AGNT Integration Path (Beyond Chat)
Annie can call Hermes from chat using the helper script. To make Hermes callable from inside AGNT workflows, Goals, and scheduled jobs, build the hermes-subagent AGNT plugin — its execute() body wraps the same delegate.py invocation. The plugin scaffold was drafted earlier; activate agnt-plugin-builder skill when ready to ship it.
