The Autonomous AI Framework
This is a working handbook on autonomous AI. How to build agents that run on their own. How to wire tools to models through the Model Context Protocol. How to design safety boundaries. How to actually ship.
Forty-two pages, organized by topic. Start anywhere. If you're brand new, start with "What is autonomous AI".
Foundations
- What is autonomous AIThe four layers that make autonomy possible→
- The autonomy spectrumChatbot → assistant → agent → autonomous→
- GlossaryEvery term, explained plainly→
Claude (the model layer)
- Claude overviewWhat Claude is, model families, why it matters for agents→
- Prompting for agentsPatterns for agentic work, not chatbots→
- Extended thinkingReasoning budgets, when they help, when they don't→
- Tool useJSON schemas, errors, multi-turn loops→
MCP. Model Context Protocol
- What is MCPThe USB-C port of AI, the standard for connecting tools to models→
- Anatomy of an MCPTools, resources, prompts, servers, clients→
- Transportsstdio, SSE, HTTP, when to use each→
- Building MCP serversServer lifecycle, patterns, examples→
- MCP clientsWho speaks MCP. Claude Code, IDEs, custom agents→
- MCP securityPermissions, sandboxing, prompt injection defense→
- MCP directoryMCPs worth trying, organized by category→
Plugins
- Plugins overviewWhat plugins are, when to use them→
- Plugins vs MCPThe difference, with a decision tree→
- MarketplacesWhere plugins come from, how to vet them→
Claude Code, the harness
- Claude Code overviewWhy it's the canonical agent harness→
- Settingssettings.json, the whole config model→
- PermissionsAllow/deny lists, auto mode, scope design→
- HooksEvent-driven automation inside the agent→
- SkillsNamed, invokable bundles of instructions→
- MemoryPersistent memory across sessions→
Agent patterns
- Patterns overview5 patterns every agent builder should know→
- ReAct loopReason → act → observe, the default→
- Plan-executePlan first, then run→
- Multi-agent orchestrationSub-agents, delegation, when it backfires→
- EvaluationHow to actually test an agent→
- Prompt cachingCost and latency wins→
Going autonomous
- What "fully autonomous" meansThe autonomy stack→
- Headless modeAgents without a human terminal→
- SchedulingCrons, triggers, event-driven→
- Self-monitoringAgents watching themselves→
- Safety boundariesGuardrails, kill switches, approvals→
Build guides
- Your first agent in 60 minutesHands-on walkthrough→
- Voice agentSTT → LLM → TTS pipeline→
- Research agentSearch → extract → synthesize → report→