Dynamic Model Delegation

Per-task model selection for Hermes Agent subagents — route coding to Claude, research to Gemini, quick lookups to Flash, all in one parallel batch.

delegate_task can't select different models per subagent. This skill orchestrates independent hermes chat -q processes, each with its own -m flag, using asyncio.TaskGroup + asyncio.Semaphore(3) for true parallel execution with structured concurrency.

What It Solves

	delegate_task	dynamic-model-delegation
Per-task model selection	❌ Single model for all subagents	✅ Different model per task
Parallel execution	✅ Built-in	✅ asyncio.TaskGroup
Structured concurrency	❌	✅ Fail-fast, cancel siblings
Zombie process prevention	❌	✅ terminate → wait → kill cascade
Characteristic-based routing	❌	✅ coding→Claude, research→Gemini, etc.

How It Works

tasks: [{goal, characteristic}, {goal, model}, ...]
         │
         ▼
    orchestrator (asyncio.TaskGroup + Semaphore(3))
         │
    ┌────┼────┐
    ▼    ▼    ▼
 hermes hermes hermes     ← each with its own -m flag
 -m X   -m Y   -m Z
         │
         ▼
    JSON results: [{goal, model, success, result}, ...]

Engine selection:

• asyncio available + >1 task → parallel (TaskGroup + Semaphore)
• Otherwise → sequential (subprocess.run fallback)
• Override: HERMES_ORCHESTRATOR_MODE=sequential

Model Mapping

characteristic	model	best for
coding	openrouter/anthropic/claude-sonnet-4	Code generation, PR review, architecture
research	google/gemini-2.5-flash	Web research, summarization, long-context
quick	deepseek/deepseek-v4-flash	Fast lookups, syntax checks
reasoning	deepseek/deepseek-v4-pro	Complex reasoning, planning
creative	openrouter/anthropic/claude-sonnet-4	Design, creative writing

Characteristic is inferred from goal keywords when not explicit (e.g., "implement" → coding, "research" → research).

Two Operating Modes

Mode 1 — Background (fire-and-forget): Spawn subagents via terminal(background=True). Results arrive asynchronously. Best for long-running tasks.

Mode 2 — Synchronous (execute_code): Call scripts/orchestrate.py with JSON tasks. Returns compiled results inline. Best for quick batches.

Usage

# Pass tasks as JSON arg
python3 scripts/orchestrate.py '[{"goal":"What is the capital of France?","characteristic":"quick"},{"goal":"Explain recursion","characteristic":"reasoning"}]'

# Or via stdin
echo '[{"goal":"...","characteristic":"coding"}]' | python3 scripts/orchestrate.py

Output:

{
  "summary": "2/2 tasks succeeded",
  "results": [
    {
      "goal": "What is the capital of France?",
      "model": "deepseek/deepseek-v4-flash",
      "success": true,
      "exit_code": 0,
      "result": "Paris."
    }
  ]
}

Installation

# Clone into Hermes skills directory
git clone https://github.com/<your-username>/dynamic-model-delegation.git \
  ~/.hermes/skills/autonomous-ai-agents/dynamic-model-delegation

# Start a new Hermes session (or run /reload-skills in an existing session)
hermes

Then load the skill: /skill dynamic-model-delegation

Verify it works:

python3 ~/.hermes/skills/autonomous-ai-agents/dynamic-model-delegation/scripts/orchestrate.py \
  '[{"goal":"Reply with exactly: pong","characteristic":"quick"}]'
# Expected: {"summary": "1/1 tasks succeeded", ...}

Prerequisites

• Hermes Agent v0.14.0+ — check with hermes --version

• Python 3.11+ — check with python3 --version

• API keys for your chosen model providers in ~/.hermes/.env. The default model map uses:

  • DEEPSEEK_API_KEY (for quick + reasoning characteristics)
  • OPENROUTER_API_KEY (for coding + creative)
  • GOOGLE_API_KEY or GEMINI_API_KEY (for research)

  You are not locked into these providers. Edit references/model-mapping.md and the MODEL_MAP dict in scripts/orchestrate.py to use any provider supported by Hermes Agent. See hermes model for available options.

• Test dependencies (only needed to run tests): pip install pytest

Repository Structure

dynamic-model-delegation/
├── SKILL.md                    # Skill definition (Hermes frontmatter + full docs)
├── README.md                   # This file
├── LICENSE                     # MIT
├── CHANGELOG.md                # Keep a Changelog format
├── CONTRIBUTING.md             # Contribution guidelines
├── SECURITY.md                 # Security policy and defenses
├── scripts/
│   ├── orchestrate.py          # Dual-engine orchestrator
│   ├── test_orchestrate.py     # 29 unit + 7 integration tests
│   └── pytest.ini              # Marker registration
└── references/
    ├── model-mapping.md        # Characteristic → model table (user-editable)
    ├── asyncio-research.md     # Asyncio design rationale with source citations
    └── multi-model-review-pattern.md  # 4-reviewer adversarial review template

Requirements

• Python 3.11+
• Hermes Agent v0.14.0+
• API keys for target model providers in ~/.hermes/.env

Tests

# Install test dependencies
pip install pytest

# Unit tests (fast, no subprocess)
python3 -m pytest scripts/test_orchestrate.py -v -k "not integration"

# Integration tests (requires hermes binary, slower)
python3 -m pytest scripts/test_orchestrate.py -v -k "integration" -n 0

# Full suite (-n 0 disables xdist — required because tests spawn subprocesses)
python3 -m pytest scripts/test_orchestrate.py -v -n 0

License

MIT — see LICENSE.