AI engineering,without the incense.
A visual field guide to the machinery wrapped around modern models. Precise enough to build with. Weird enough to remember.
The machine is already strange. We do not need to make it mystical.
Learn the system.Keep the wonder.
Most AI explanations choose between hand-waving and a wall of API reference. These notes build an internal movie first, then show the moving parts, failure modes, and practical receipts.
Choose your particular flavor of confused.
Read in order for a gentle ramp, or jump straight to the thing that broke in production five minutes ago.
Make the model useful.
Start with repeated work, limited memory, meaning as coordinates, and evidence retrieved just in time.
Turn answers into work.
Follow tool calls into loops, coding workflows, harnesses, and standard connections to the outside world.
Know what happened.
Measure behavior, constrain consequences, and trace the exact path from request to weirdness.
Every note is open.
The plumbing
4Agent behavior
5Tool calls
How models ask your code to do things in the real world.
Agent loops
The observe, decide, act, inspect rhythm under the robot costume.
Coding agents
How an agent navigates a repo, edits files, and checks its work.
Harnesses
The runtime that gives a model tools, memory, boundaries, and a shift.
MCP
A shared socket shape for models, tools, and context.
Make it good
3The vocabulary will change. The mechanics leave fingerprints.
Providers rename things. Models get more capable. Good mental models still help you ask: what entered context, who chose the action, what executed it, what boundary applied, and what evidence says it worked?
Inspect the rig around the model