Skip to content
Longread and sources
Giga Conf · October 16, 2026Alexander Polomodov

Harness: past, present and future

The layer between the model and the world: what to simplify after a new model, and what to reinforce

/ Harness · Giga Conf · October 16, 2026

Slide contents

  1. 1. Harness: past, present and future

    The layer between the model and the world: what to simplify after a new model, and what to reinforce

  2. 2. Six steps through the harness

  3. 3. Words & walls

  4. 4. The harness: all but the weights

  5. 5. 01. How the harness emerged

    From a loop in the prompt to a loop as a service

  6. 6. Five eras in four years

  7. 7. The word “harness” came last

  8. 8. Three one-way migrations

  9. 9. 02. What it is made of

    The surface converged; the defaults differ

  10. 10. Everyone shares one vocabulary

  11. 11. Protocols split along axes

  12. 12. The API line is your state

  13. 13. Four harnesses split the market

  14. 14. Who stops a risky step differs

  15. 15. Where the loop runs, whose model

  16. 16. 03. Combine, agent, block

    Three ways to look at one layer

  17. 17. Three views of one harness

  18. 18. Everyone converges on the combine

  19. 19. Hypothesis: pure agents go niche

  20. 20. The block is a three-step ladder

  21. 21. Bash in 100 lines maxed SWE-bench

  22. 22. 04. Never trust models with permissions

    Not even the strongest: why walls are code

  23. 23. Never delegate authorization

  24. 24. 0.5% per attempt is a lot

  25. 25. Code, not prompts, closes holes

  26. 26. Classifiers sit inside the walls

  27. 27. 05. How to prove value

    Points, repeatability and the cost of an accepted result

  28. 28. Harness effect: two regimes

  29. 29. Words shrink, environment stays

  30. 30. Pass@1 alone is not enough

  31. 31. One branch, one component

  32. 32. Ledger: assumption and trigger

  33. 33. Cheaper tokens, pricier work

  34. 34. Context and cache move the cost

  35. 35. Output, time and understanding

  36. 36. 06. What comes next

    Agents with their own machines, checks at the vendor

  37. 37. Agents get their own machine

  38. 38. Vendors take over action checks

  39. 39. Three testable bets

  40. 40. What to do after the talk

    Keep a harness assumption ledger

    Retest and remove compensating instructions

    Walls in code, classifiers inside

    Measure repeatability and cost per success

    Prove each component by ablation

    Learn to simplify your harness

  41. 41. Thank you!

    AI4SDLC

    Slides, the harness longread and sources are at polomodov.tech; AI4SDLC news is in the Telegram channel

    Alexander Polomodov, CTO & Technical Fellow, polomodov.tech

    @ai4sdlc