Local coding agent with Hermes and Ollama

• Understand the role of Hermes, Ollama, and the local model.
• Configure a safe workflow for editing code on a branch.
• Avoid loops, broken tool calling, and changes without tests.

Minimal architecture

For a local coding agent, think in three layers: Hermes orchestrates the task, Ollama serves the model, and Git/tests verify whether the change deserves to stay. If one of those layers is missing, the system becomes fragile.

• Hermes: memory, skills, tools, and work loop.
• Ollama: compatible local endpoint for inference.
• Repository: isolated branch, tests, lint, and reviewable diff.

# Check that Ollama is responding
ollama --version
ollama list
curl http://localhost:11434/api/tags

# Choose a small model to start with
ollama pull qwen2.5-coder:7b
ollama run qwen2.5-coder:7b "Summarize what this package.json file does"

# Always work on a branch
git checkout -b agent/hermes-prueba-segura

Recommended initial assignment

A local agent should not start by editing. It should observe, scope, and propose a plan first. This pattern reduces impulsive changes and lets you detect when context is insufficient.

Objective:
Fix the bug described in BUG.md without touching other areas.

Rules:
- First list relevant files and risks.
- Do not edit until you propose a 5-step plan.
- Maximum 8 agent steps.
- If a tool fails twice, stop and summarize.
- After editing, run available lint/test/build.
- Do not commit without showing the diff and test results.

Stop logic so it doesn't wander off

limits:
  max_steps: 8
  max_tool_calls_total: 20
  max_repeated_tool_args: 2
  max_runtime_minutes: 15
  stop_when:
    - tests_pass
    - same_error_twice
    - missing_context
    - needs_secret_or_external_login

Official sources

• Ollama: Hermes Agent integration
• Hermes Agent documentation
• Hermes Agent: AI providers
• Google Gemma 4 model card
• Ollama API