The Local Agent Playbook — Safe, Reliable AI Agents on Your Own Hardware

Local models are good enough to run agents now. Keeping those agents useful and safe is the hard part.

This playbook collects proven, widely used agent-engineering patterns into a form you can copy straight into your own projects:

• Evidence-first prompting — agents that say "not verified" instead of inventing facts
• Approval gates enforced in code — nothing is sent, posted, deleted or bought without a human seeing the exact action
• Reliable tool calls — tight JSON Schemas + a repair loop that fixes most broken calls
• Loop breakers — novelty checks, step budgets and stuck detection (no more 40 identical plans overnight)
• Decision-log memory — facts, decisions and open loops instead of dumping the whole chat history
• Operations — heartbeats, health checks and daily summaries

Included

• 10-section PDF playbook (+ Markdown)
• 5 copy-ready prompts: evidence-first system prompt, planner, critic, summariser, fact-bound writer
• mini_agent.py — ~200-line reference agent, Python standard library only: OpenAI-compatible tool calling, risk-classified tools, approval gate, repair loop, audit log
• Works with llama.cpp, LM Studio, Ollama (OpenAI endpoint) and vLLM

Not included: model weights, a hosted service, or integrations with your accounts (the demo "send" tool only prints). Parts of the text were drafted with AI assistance and edited and tested by hand.