Skip to content
Written summaryAddy Osmani · primary postHuaShu · working note
Research Insights Made Simple logo
Research Insights Made Simple #19 · July 14, 2026Podcast · with Maxim Smirnov

Loop Engineering

How to design systems that launch coding agents themselves—and remain controllable

/ Research Insights Made Simple #19 · Loop Engineering

Slide contents

  1. 1. Loop Engineering

    How to design systems that launch coding agents themselves—and remain controllable

  2. 2. IEEE styling does not confer IEEE status

    What it is

    A 2026 working note

    A reformatted Orange Book

    A practical playbook

    What it is not

    Not an IEEE publication

    Not an Anthropic paper

    Not a benchmark study

  3. 3. Engineers move from execution to loop design

    Inside the loop

    Give the next prompt

    Watch every turn

    Start work manually

    Outside the loop

    Define discovery

    Design the evaluator

    Keep stop points

  4. 4. The practice was named within one week

  5. 5. Loop sits one floor above the harness

    Redrawn from HuaShu, Fig. 1 and Table I · PDF p. 2

  6. 6. Higher layers let errors survive longer

  7. 7. One turn passes through five required moves

    Redrawn from HuaShu, Fig. 2 and Table II · PDF p. 3

  8. 8. Morning triage turns signals into reviewable PRs

    Redrawn from HuaShu, Table II · PDF p. 3

  9. 9. Six parts implement five moves

    Redrawn from HuaShu, Table III · PDF p. 4

  10. 10. Repeated runs do not make a loop

  11. 11. The hardest component must say no

    Verify results, not confident self-reports

    Treat reject as a normal outcome

    Feed reasons into the next turn

    Green tests do not mean right

  12. 12. Maker and checker must stay independent

    Redrawn from HuaShu, Fig. 3 · PDF p. 4

  13. 13. The evaluator should act like a user

  14. 14. Full harness runs cost far more

    Retro game · Opus 4.5

    Solo · 20 min · $9

    Full harness · 6 h · $200

    Same prompt, different output

    DAW · V2 harness

    3 h 50 min

    $124.70

    Opus 4.6 · different prompt

  15. 15. The evaluator pays off beyond reliable solo

  16. 16. Every skipped move creates an anti-pattern

    Redrawn from HuaShu, Fig. 4 · PDF p. 5

  17. 17. Deterministic gates surround LLM work

    Redrawn from HuaShu, Fig. 5 · PDF p. 6

  18. 18. Scale holds; exact numbers depend on source

    HuaShu · podcast retelling

    1,300+ PRs / week

    Wording: merged

    Machine-written

    Stripe Sessions · official

    Over 1,000 / week

    Shipped to production

    Human review + approval

  19. 19. Three scheduler modes carry different trust budgets

    Redrawn from HuaShu, Table IV · PDF p. 6 · June 2026 product snapshot

  20. 20. Four debts reinforce one another

    Redrawn from HuaShu, Fig. 6 · PDF p. 7

  21. 21. Green PRs can hide a control failure

  22. 22. Discovery turns external text into control input

  23. 23. Autonomy requires bounded, reversible actions

  24. 24. Cheap generation makes judgment the bottleneck

  25. 25. Three disciplines preserve the right to refuse

  26. 26. Start with a small, complete loop

    Redrawn from HuaShu, Table VI · PDF p. 9

  27. 27. The playbook hypothesizes; our eval supplies evidence

  28. 28. A loop should run itself— and remain stoppable

    Book Cube

    Working note, primary-source links, and new engineering research breakdowns are in the channel

    Alexander Polomodov, Technical Director & Fellow, T-Technologies

    @Book_Cube