All content
264 pages · 5 levels
Concepts
- What is an AI agent?
- Building agents in ChatGPT - so why a platform?
- Block I · Agent harness
- The LLM Call
- Message History
- System Prompt
- Tools & Loop
- Going deeper
- Deep dive
- Try it interactively
- When does the LLM pick a tool?
- Tool descriptions are decisive
- Anatomy of a tool_use block
- Parallel tool calls
- Direct answer vs. tool call
- The agent loop step by step
- The messages array after one roundtrip
- Phase 4 vs. Phase 5 - recognition vs. execution
- Routing - five questions, five right picks
- Multi-step task - sequence from the model
- The loop is agnostic - 2 vs. 5 tools
- Parallel tool calls - one iteration, many calls
- Tool shyness - measure the hit rate, tune the prompt
- Quiz
- Quiz
- Quiz
- Guardrails
- Block II · Extensions
- Memory
- MCP servers
- Sidecars
- Tool calling in production
- Tiny Agents and the Compression Boundary
- Guardrails in production
- The damage catalogue
- The map: where, when, what, how
- The gatekeeper: build forms and data path
- The detectors: four families and their real numbers
- The entry gate: normalise, mask, mark
- The exit gate: the last look before delivery
- The handbrake: hard rules at the action level
- The limits: what the attack challenges really show
- Channels & triggers
- Swarm
- Block III · Inference
- How do the others do it?
Tutorials
- Get an OpenRouter key and make your first call
- How neural networks learn to read - the MNIST lab
- A Soul in a File - what belongs in a soul.md (and what doesn't)
- Hugging Face for beginners
- Understanding RAG - step by step with real data
- Build your own PyTorch
- Build your own tiny agents system
- Build an MCP server that does something
- Fine-tune FunctionGemma on your own MCP
Explore
- Aha!
- What is behind System One and the open Jev clones?
- The world, or a model of the world
- Why the same question takes 150 seconds one time and 4 the next
- Why your 8 GB model needs 20 GB of memory
- Why the tenth turn costs a hundred times what the first one did
- Why a model sounds more certain when it's wrong than when it's right
- Why the answer stops mid-sentence
- Repos
- News
- Community
- Arena
- Training ground
- Input-format lab
- AI driving licence - the exam
- AI driving licence - the simulation
- Benchmarks
- Changelog
Library
- The workbook
- Tool calling in production
- How an LLM works
- Tiny Agents
- The Enterprise Harness
- Guardrails in production
- Memory in production
- The little harness
- The teacher handbook
- The AI driving licence - trainer's guide
- The AI driving licence - participant booklet
- gemma-4-31B-it
- Qwen3.6-35B-A3B
- Qwen3.8-27B
- Qwen vs. Qwen