Skip to content

Foundations

Mental Models

Three pictures to carry in your head for the whole course: the operator you're replacing, the two loops that unfortunately share one name, and the body…

Three pictures to carry in your head for the whole course: the operator you're replacing, the two loops that unfortunately share one name, and the body a loop is built like.

The hook

For roughly two years, "using a coding agent" meant nursing the tool one turn at a time: write a prompt, read the result, type the next thing. Now zoom out and watch the person doing that all day — finding the work, handing it out, checking it, recording what happened, deciding what comes next. Loop engineering replaces that operator with a small system you design once. The agent didn't suddenly get smarter. The management around it did.

Model 1 — you move up one seat (plain English)

Here's the reassuring part: your value doesn't evaporate when the operator is automated. It concentrates — at the two ends a loop can never own for you.

  • Intent — stating what you want precisely enough that the result can be checked.
  • Accountability — standing behind whatever ships.

Everything in between — triggering, doing, verifying, logging — is now the loop's job, not yours.

Model 2 — two loops share the name

Diagram

The small/inner loop is the agent's own context → tool calls → results → repeat cycle. Its only native way to stop is the model's own self-assessment — which is the source of every confident-but-wrong "Done!" you have ever been handed. The big/outer loop is the manager wrapped around it: it chooses the task, sets the timing, does the grading, keeps the memory. Hold onto this one line: one inner-loop run is exactly one beat of the outer loop.

Model 3 — the loop as a body

Body partLoop partJob
HeartHeartbeatstarts each beat (schedule, event, condition)
SpineState filedurable memory so runs compound, not restart
HandsConnectors (MCP)act on the world, not just suggest
Immune systemCheckergrades output; maker never grades itself
SkeletonWorktree/isolationparallel work that can't collide
Trained reflexesSkillsproject knowledge written once, read each run

[!NOTE] Going deeper: the body metaphor comes from Addy Osmani's Loop Engineering (S5), and the six-part anatomy from the Panaversity backbone (S1). Part 1 (Day 2) takes each part apart in depth.

Check yourself

Q: Your agent announces "All tests pass, task complete!" — which loop produced that claim, and why shouldn't the outer loop take it at face value?

Answer

The inner loop: it stopped on the model's self-assessment. The outer loop treats that as a signal to verify, never as a verdict — its checker runs the tests itself (maker ≠ checker). Green from the maker is a claim. Green from the checker is a fact.

Try With AI

Ask your agent:

"Walk me through your own loop: what starts your turn, what do you see, what tools can you call, and what makes you decide you're finished?"

Now map its answer onto the inner-loop diagram above. Pay special attention to the last part of its reply — how it decides it's finished. That single gap is the entire reason the outer loop exists.

When it goes wrong

SymptomCauseFix
"The agent said done but it isn't"You trusted the inner loop's self-stopAdd an outer-loop checker with a machine-checkable stop
Same mistake every morningNo spine — each run starts amnesiacAdd a state file; read it first, update it every beat
You babysit every run anywayIntent was never made checkableRewrite the goal as a spec (see the spec-driven primer)

Glossary terms used on this page: beat, spine, checker, heartbeat — see glossary.md.

Sources: the two-loop and body metaphors come from Addy Osmani's Loop Engineering (S5) and the six-part anatomy of Panaversity's Loop Engineering: A Crash Course (S1). Full attribution: resources/sources.md.