Don't Waste Your Backpressure: Engineering Better AI Harnesses as a Team

An AI coding agent can produce a convincing change and still leave your
team explaining the same conventions, checking the same things, and catching
the same mistakes. One engineer adds a prompt, another adds a rule, and the
next session starts the argument again. Everyone has a favourite setup.
The team still has the same problem. We are giving agents plenty of
feedback, but how much of it helps anyone beyond the person typing it?

From a forward-deployed engineer's perspective, I'll explore how teams can
build and improve a shared engineering harness: the tools, context, and
controls around their coding agents. We'll look at how to manufacture useful
backpressure so agents can correct their work before it reaches a human,
then use telemetry to find where that feedback is helping and where it is
creating expensive friction. I'll show how to refine noisy sensors, move
repeatable checks from model judgment into tools, and reduce token waste
without quietly handing more work back to reviewers. You'll leave with a
practical improvement cycle for your team: find a recurring failure, improve
the harness, measure the result, and share what works. The goal isn't more
rules. It's to stop paying for the same lesson twice.

Transcript

1. Don't Waste Your Backpressure Engineering Better AI Harnesses as a Team Dasith Wijesiriwardena @dasiths Australia 28th October 2026
2. HELLO! I’m Dasith dasiths dasith.me
3. https://aauth.dev Nanasa (නැනස) is a localfirst orchestrator for running and observing multiple codingagent terminals. dasiths dasith.me
4. Agenda Why should you care? Backpressure in AI assisted engineering Making use of friction to improve your harness Improvement loop @dasiths
5. Chapter 1 Why should you care? @dasiths
6. @dasiths
7. @dasiths
8. Alignment Gap @dasiths
9. More code is not the same as more progress @dasiths
10. Alignment Gap Feed-forward @dasiths
11. Feed-forward - Specs PRD/BRD Architecture.md Style guide Alignment Gap @dasiths
12. Feedback Alignment Gap @dasiths
13. @dasiths
14. Small tasks, more frequent reviews @dasiths
15. Your attention is valuable ▪ Your attention can become the bottleneck ▪ Unchecked mistakes can become foundations ▪ Repeated corrections are a team-level opportunity @dasiths
16. Opportunity Can we improve the shared harness, so the next person doesn’t have to explain it again? @dasiths
17. Chapter 2 Backpressure @dasiths
18. Backpressure Feedback that constrains an AI agent’s progress, requiring it to address unmet expectations before its work can be accepted. @dasiths
19. Human only backpressure @dasiths
20. Human only backpressure @dasiths
21. Sensor A mechanism that checks an agent’s work against defined expectations and reports evidence of where it meets or falls short of them. @dasiths
22. Gate Sensors detect unmet expectations. Gates turn those findings into backpressure, requiring the agent to address them before proceeding. @dasiths
23. Example ▪ Sensor: A coverage tool reports that tests execute 65% of the changed lines. ▪ Gate: The team requires at least 80% coverage of changed lines before merging. ▪ Backpressure: The change is blocked, requiring the agent to add tests and rerun the checks rather than declare the work complete. @dasiths
24. Sensors Sensors+ +Gates Gates @dasiths
25. Sensors + Gates @dasiths
26. Example Let’s create a sensor @dasiths
27. Enforce an Architecture Imagine a team working in a layered .NET codebase. The layers are … ▪ ▪ ▪ ▪ Shop.Api Shop.Application Shop.Domain Shop.Infrastructure @dasiths
28. Enforce an Architecture ▪ ▪ ▪ ▪ ▪ Controllers in Api do not talk to repositories in Infrastructure directly. Domain does not depend on Application, Api, or Infrastructure. Command handlers live in Application. DTOs stay in Api. Domain stays persistence-ignorant: no infrastructure or ORM types, and no persistence assumptions leaking through method shapes or naming. @dasiths
29. Reviewer Sub-Agent You are reviewing a change for architectural fit in a layered codebase. The layers, outermost to innermost, are Api, Application, Domain, Infrastructure. Check the change against these rules: - Controllers in Api must not depend on repositories in Infrastructure directly. - Domain must not depend on Application, Api, or Infrastructure. - Command handlers must live in Application. - DTOs must stay in Api. - Domain must not reference any infrastructure or ORM type. - Each type sits in the layer its responsibility belongs to, not just a legal one. - Abstractions are worth their weight, not just ceremony. - No lower-layer assumption leaks through method shapes or naming without a direct dependency. - The domain is modelled correctly, with the right invariants and value objects for the business rule. For each problem, name the offending type and the rule it breaks. @dasiths
30. Reviewer Sub-Agent You are reviewing a change for architectural fit in a layered codebase. The layers, outermost to innermost, are Api, Application, Domain, Infrastructure. Check the change against these rules: - Controllers in Api must not depend on repositories in Infrastructure directly. - Domain must not depend on Application, Api, or Infrastructure. - Command handlers must live in Application. - DTOs must stay in Api. - Domain must not reference any infrastructure or ORM type. - Each type sits in the layer its responsibility belongs to, not just a legal one. - Abstractions are worth their weight, not just ceremony. - No lower-layer assumption leaks through method shapes or naming without a direct dependency. - The domain is modelled correctly, with the right invariants and value objects for the business rule. For each problem, name the offending type and the rule it breaks. Inferential Sensor @dasiths
31. Split The Rubric Some conditions are structural facts ▪ ▪ Controllers must not depend on repositories DTOs stay in Api ▪ Does this abstraction help, or is it only ceremony Other conditions are semantic judgments @dasiths
32. Architecture Fitness Tests NetArchTest @dasiths
33. Encode Your Rules ▪ Strongly typed languages help ▪ Remove footguns ▪ Deterministic rules are cheaper than inferential ones @dasiths
34. A Better Inferential Sensor Reviewer Sub-agent Prompt @dasiths
35. A Better Inferential Sensor Reviewer Sub-agent Prompt @dasiths
36. Deterministic + Inferential Sensors We still have an inferential sensor. It is now a better one. We rely on the deterministic sensors for encoded checks and the inferential sensors for the residue. @dasiths
37. Chapter 3 Friction @dasiths
38. Opportunity Every recurring struggle is a candidate for a shared tool, instruction, or better feedback mechanism. @dasiths
39. Friction Friction is the effort, delay, or resistance encountered while turning an intention into acceptable work. @dasiths
40. Two Kinds ▪ Protective friction: a test rejects incorrect behaviour, or a gate requires human approval. It exists for a reason. ▪ Avoidable friction: the agent repeatedly guesses command syntax, reconstructs repository conventions, or retries without understanding an error. @dasiths
41. Observability Where does the agent repeatedly spend effort without making useful progress? @dasiths
42. Observability Where does the agent repeatedly spend effort without making useful progress? @dasiths
43. Example https://github.com/dasiths/github_copilot_grafana_extensions @dasiths
44. Example @dasiths
45. Example Tool failures are an easy opportunity @dasiths
46. Improving Your Harness Your coding agent harness is the foundation. Build on top of and make it your own “engineering harness”. Use a project specific CLI to do common tasks… ▪ ▪ ▪ Read backlog, conventional branches and work tracking Running sensors and reporting back Wrap failing tools Make this CLI available to your coding agent as a first-class surface. @dasiths
47. AXI: Agent eXperience Interface https://github.com/kunchenguid/axi @dasiths
48. Example: Custom Harness https://github.com/dasiths/engineering-harness-axi @dasiths
49. Chapter 4 The Loop @dasiths
50. Continuous Improvement Don’t just help this agent get unstuck. Improve the shared environment so the next agent doesn’t get stuck in the same place. @dasiths
51. @dasiths
52. Don’t just correct the agent. Improve the harness. Turn recurring friction into shared tools, guidance, and checks for the whole team. Don’t waste your backpressure @dasiths
53. https://www.nationalgeographic.com/travel/destinations/asia/sri-lanka/ THANKS! Any questions? @dasiths dasith.me
54. CREDITS Special thanks to all people who made and shared these awesome resources for free: Presentation template designed by powerpointify.com Photographs by unsplash.com Free Fonts used: https://www.fontsquirrel.com/fonts/oswald @dasiths
55. Colour Palette INSPIRED BY SRI LANKAN TRADITIONAL ART Primary Red Accent Red Terracotta Ochre Mustard Cream Parchment Leaf Green Deep Green Earth Brown Charcoal Stone Grey Positive / Success Secondary Text #9E1B1B #E8D8B0 #C62828 #4E6B46 #D5522F #D4A017 #2F4F3E #E0B12F #5B3A24 #1F1F1F #F1E3C6 #8E8E8E USAGE SUGGESTIONS Section Label #C62828 Primary Text #F1E3C6 Highlight / CTA #D4A017 #2F4F3E #8E8E8E