Hermes Agent Optimization

Source Excerpt

Domain: AI Agents, Fine-Tuning, Tool Calling

Status: Active Research

Strategic Importance: High — Core to retired internal project agent capabilities

Executive Summary

Hermes Agent Optimization is a specialized approach to training language models for tool-calling and multi-step task execution within the Hermes Agent framework. Unlike generic chat models or benchmark-optimized systems, Hermes-optimized models prioritize harness-native behavior — executing tools accurately, handling terminal operations, managing multi-turn workflows, and maintaining proper action structure inside the Hermes runtime.

Key Insight: Models trained on Hermes-specific traces (Carnice Model Family-9b) outperform generic tool-trained models because they learn the actual patterns and constraints of the harness, not abstract tool schemas.

Core Principles

1. Harness-Native Over Generic

Problem: Most "agent" models are trained on generic tool schemas (OpenAI Functions, generic JSON schemas) that don't match the actual structure and flow of real agent frameworks.

Solution: Train on actual Hermes traces — real conversations, tool calls, terminal outputs, file edits, and multi-step workflows from inside the Hermes harness.

Result: Models that "think in Hermes" — natural action structure, proper error handling, correct tool sequencing.

2. Execution Quality Over Benchmark Scores

Problem: Leaderboard optimization (MMLU, GSM8K, HumanEval) doesn't correlate with agent execution quality.

Solution: Optimize for:

• Tool calling accuracy (did the model call the right tool with correct arguments?)
• Terminal behavior (does it handle shell output, errors, long-running processes?)
• File editing correctness (proper diffs, atomic changes, rollback capability)
• Multi-turn workflow completion (can it execute 10+ step tasks without losing context?)

Result: Models that actually get work done instead of looking good on benchmarks.

3. Two-Stage Training Architecture

Stage A — Reasoning Repair:

• High-signal reasoning datasets (Bespoke-Stratos-17k, NuminaMath-CoT)
• Builds foundational reasoning capability
• Teaches chain-of-thought, problem decomposition

Stage B — Hermes-Specific Refresh:

• Harness-native traces (Carnice Model Family-glm5-hermes-traces, OpenThoughts-Agent-v1-SFT)
• Teaches Hermes action structure
• Reinforces tool calling patterns specific to Hermes runtime

Critical: Stage B is the differentiator — without it, models have generic reasoning but awkward harness behavior.

Technical Implementation

Training Data Requirements

Model Architecture Considerations

Base Model Selection:

• Qwen3.5-9B (Carnice Model Family-9b base) — good reasoning, efficient
• Qwen3.5-27B (pending) — better reasoning, still tractable
• Avoid overly large models (70B+) unless necessary — diminishing returns for agent work

Precision:

• Training: bfloat16 (stability)
• Deployment: Q4_K_M to Q6_K GGUF Format Format (balance quality/size)

Context Length:

• Minimum: 8K (tool outputs, file contents)
• Target: 16K-32K (complex workflows, multi-file edits)

Evaluation Metrics

Not These (Generic Benchmarks):

• MMLU score
• GSM8K accuracy
• HumanEval pass@1

These (Agent-Specific):

• Tool call accuracy (% correct tool + arguments)
• Workflow completion rate (% multi-step tasks finished)
• Error recovery success (% failures handled correctly)
• Terminal behavior quality (proper output parsing, timeout handling)
• File edit correctness (% atomic, reversible changes)

retired internal project Applications

Current Implementation

Carnice Model Family-9b is the primary Hermes-optimized model for retired internal project:

Performance Characteristics:

• ✅ Excellent tool calling accuracy (Hermes-native training)
• ✅ Proper terminal behavior (shell output parsing, error handling)
• ✅ Correct action structure (Hermes format, not generic JSON)
• ✅ Multi-turn workflow capability (10+ step tasks)
• ⚠️ 9B size limits complex reasoning (27B pending)

Deployment Strategy

Mac Mini (16GB) — Edge Nodes:

Model: Carnice Model Family-9b @ Q4_K_M (~6GB)
Context: 8K-16K
Use: Small workspaces, single-agent tasks
Concurrent: 1-2 instances

Mac Studio (256GB) — Central Hub:

Model: Carnice Model Family-9b @ Q8 or bf16 (~18-20GB)
Context: 16K-32K
Use: Primary agent, complex workflows
Concurrent: 2-3 instances (with larger model)

Ubuntu 3090 (24GB) — GPU Workers:

Model: Carnice Model Family-9b @ bf16 or Q6_K (~18GB)
Context: 16K-32K
Use: Batch inference, fine-tuning experiments
Concurrent: Multiple parallel workspaces

Future: Carnice-27B

Expected Benefits:

• Better reasoning depth (27B vs 9B)
• More complex workflow handling
• Improved error recovery
• Still fits Mac Studio (Q4_K_M ~18GB) and 3090 (Q4_K_S ~20GB)

Action Plan:

1. Monitor @kai-os HuggingFace for release
2. Immediate download and validation
3. Quantize to Q4_K_M, Q5_K_M, Q6_K
4. Benchmark vs 9B on retired internal project workflows
5. Deploy as primary model if improvement confirmed

Comparison to Alternatives

Key Researchers & Sources

Related Concepts

• concepts/hermes-agent-optimization — Broader methodology
• concepts/tool-calling-patterns — Specific action structures
• ai-ml/fine-tuning-methodologies — Training approaches
• retired-internal-project/agent-architecture — System design
• concepts/nightly-learning-loop — retired internal project's self-improvement mechanism

Open Questions

1. Optimal Context Length: Is 16K sufficient, or does 32K meaningfully improve complex workflows?
2. Fine-Tuning Depth: Should retired internal project fine-tune Carnice Model Family further on proprietary workflows?
3. Multi-Model Strategy: When to use 9B vs 27B vs larger models?
4. Evaluation Benchmark: What's the right retired internal project-specific agent benchmark?

Follow-up Actions

• [x] Document Hermes optimization principles
• [x] Link to Carnice Model Family-9b model page
• [ ] Create retired internal project agent benchmark suite
• [ ] Test Carnice Model Family-9b on representative workflows
• [ ] Plan Carnice Model Family-27B integration (when released)
• [ ] Evaluate fine-tuning on retired internal project-specific traces

Metadata

• Created: 2026-04-05 by Julius
• Last Updated: 2026-04-05 by Julius
• Next Review: 2026-04-12 (weekly — aligns with Carnice Model Family-27B watch)

Source excerpt truncated at 220 of 222 lines. Open the canonical wiki path above for the full page.