Agentic AI
Architect systems that plan, use tools, retain state, and coordinate work across specialized components.
ORCHESTRATION / MEMORY / CONTROLENTER SECTION →NOT WHAT. NOT WHY.
HOW?
The engineering manual
for the AI age.
Reproducible guides for building agents, engineering reliable loops, controlling context, evaluating frontier models, and running open models locally.
FIGURE 01RELIABILITY / FEEDBACK / CONTROL
THE RELIABILITY LOOP
An agent is only as good as its correction path.
Select a phase to inspect the operating logic. A useful loop does more than retry: it knows what changed, what failed, and when to stop.
01 / 04
GOAL → ACTION
Plan
Turn intent into one testable next action with an explicit success criterion.
MANDATORY CHECKSTOP CONDITION?yes → return result / no → continue within budget
CONTENTS / 01–06THE COMPLETE WORKING SYSTEM
THE FIELD MANUAL
Six disciplines. One job: make AI systems work.
Agents
Build, deploy, instrument, and evaluate agents that behave predictably beyond the demo.
TOOLS / STATE / EVALUATIONENTER SECTION →Loop Engineering
Design bounded plan → act → observe → correct loops with recovery rules, budgets, and safe stopping conditions.
RECOVERY / BUDGETS / STOPPINGENTER SECTION →Prompts
Engineer instructions, context, and tool contracts that remain stable under real inputs.
CONTEXT / CONTRACTS / OUTPUTENTER SECTION →Frontier Models
Understand how current GPT, Claude, and Gemini models behave through testing, not benchmark theater.
CAPABILITY / REASONING / TESTSENTER SECTION →Local Models
Select models, quantizations, runtimes, and hardware for private inference on your own machines.
WEIGHTS / HARDWARE / INFERENCEENTER SECTION →FEATURED FIELD GUIDE
001LOOP ENGINEERING / 18 MINHow to Engineer a Self-Correcting Agent Loop
Build a bounded agent runtime that detects failed assumptions, selects a recovery strategy, and stops safely when another attempt is unlikely to help.
- Model loop state explicitly
- Classify recoverable and terminal failures
- Set retry, token, latency, and tool budgets
- Record traces that explain every correction
PYTHONSTATE MACHINESEVALUATION
Open the field guide ↗Includes a reference loop, typed state schema, recovery matrix, stop conditions, and an evaluation harness you can adapt.
REVISION LOG6 ENTRIES SHOWN
NEW + REVISED
The latest working notes.
Specific enough to implement. Honest about uncertainty. Revised when the evidence changes.
01REV. 07.08
Prompts / 11 MIN
How Context Windows Actually Work
Token capacity is not usable attention. See how placement, repetition, retrieval, and compaction change what a model can actually use.
02REV. 07.07
Local Models / 14 MIN
How to Run a 70B Model on a MacBook
Choose a quantization that fits unified memory, configure a local runtime, then benchmark quality, latency, and tokens per second.
03REV. 07.04
Prompts / 12 MIN
How to Write System Prompts That Don’t Break
Turn vague instructions into an operating contract with authority rules, tool boundaries, output schemas, and failure behavior.
04REV. 07.02
Agentic AI / 16 MIN
How to Orchestrate Multi-Agent Handoffs
Define ownership, handoff contracts, context limits, timeouts, and escalation paths so specialized agents do not lose the task.
05REV. 06.30
Frontier Models / 13 MIN
How Frontier Labs Train Reasoning Models
A practical account of pretraining, instruction tuning, verifiable rewards, and inference-time reasoning, with evidence clearly labeled.
06REV. 06.27
Agents / 15 MIN
How to Evaluate an Agent Before You Ship It
Build a representative task suite, score outcomes and traces, test recovery paths, and gate releases against behavioral regressions.