Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Disciplined Vibes: Scaling AI-Assisted Engineer...

Disciplined Vibes: Scaling AI-Assisted Engineering (By: Sheharyar Nasser) - Build with AI 2026

Talk by Sheharyar Naseer (https://www.linkedin.com/in/sheharyarn/) at Build with AI 2026 by GDG Lahore

GDG Lahore PRO

June 17, 2026

Transcript

33 slides
  1. Background Principal Software Architect at Infra One ✦ Worked with:

    Apple, Slab, TheScore, Superlist, etc. ✦ 16+ years of polyglot experience, focus on Web & Cloud ✦ StackOver ow: 75,000+ score (Top 5 in Pakistan) ✦ Author / Contributor of multiple famous libraries & tools ✦ Featured on popular developer communities fl ✦
  2. PA R T 1 The Problem PA R T 2

    It's Not the Model PA R T 3 Harness Engineering PA R T 4 Outline Live Workshop OU T R O What's Next?
  3. Struggling with AI 16+ years of experience, still humbled by

    a chatbot ✦ Struggled a lot with AI-assisted coding ✦ Code quality was extremely poor ✦ Often had to spent time xing it ✦ Or throwing it away and doing manually fi ✦
  4. AI-Assisted Problems ✦ Hallucinated APIs, function calls, and packages ✦

    Insecure code ✦ Architectural drift ✦ Ignored edge-cases ✦ Incorrect, or no error-handling ✦ Performance issues ✦ So many more...
  5. The Data Agrees METR DORA METR's randomized controlled trial found

    experienced developers were 19% slower with early-2025 AI. Google's DORA 2024 research found AI adoption reduced delivery stability, continuing into 2025 despite higher adoption & throughput. SOURC E S O UR CE
  6. “ Seniors often get worse results than juniors from same

    tools until they learn deliberate prompting. But once they do they have a massive advantage. Sabrina Goldfarb SWE at Github Co-Pilot
  7. 02 It's Not The Model Exploring the root causes and

    developing the right thinking model.
  8. It's a You Problem ✦ Don't understand how LLMs work

    ✦ ✦ fi ✦ Gold sh memory & context management Incomplete specs ✦ Basic prompts ✦ Missing documentation & examples Unreliable guardrails ✦ No systems or quality checks ✦ Agents don't receive feedback about what's wrong
  9. Mental Models AI Search Vibe Coding Vibe Engineering Shallow use

    of modern LLMs as a Google replacement Fully delegating code to AI without reviewing output Accelerating professional software engineering with AI YO U AR E HE RE
  10. Vibe Engineering ✦ Does not mean better prompts ✦ Foundation/architecture/system

    where the agent can "succeed" ✦ Feedback loops ✦ Also called Evaluation Driven Development (EDD)
  11. “ You shouldn’t be prompting coding agents anymore. You should

    be designing loops that prompt your agents. Peter Steinberger Creator of OpenClaw, Technical Staff at OpenAI
  12. What's a Harness? ✦ LangChain research team describe it as:

    Agent = Model + Harness ✦ "Everything other than the model" ✦ Prompt, Evals, Tool Calls, Docs, Context, etc. ✦ Even the GUI/CLI "agent" tool you use “ Agent = Model + Harness Vivek Trivedi (Researcher, LangChain)
  13. The Model Doesn't Matter ✦ SWE bench score improvements ✦

    42% → 78%, 46% → 80%, 23% → 45% ✦ ~22 point swings vs ~1 point swings ✦ Using frontier models S A ME HA R N ES S Different Model S A ME MO DE L Scaffold Changes ~1 PO IN T SW IN GS ~22 PO IN T SW IN GS
  14. Anatomy of a Harness ✦ ✦ Inner Harness (System) ✦

    Built into your coding agent (CLI/GUI tool) ✦ System prompt, Tool calls, Orchestration M O DE L Outer Harness (User) I N N E R H ARN E SS ✦ Controls put in place by users ✦ User prompt, Agent rules, Output validation ✦ Our focus today O UTER HARNESS
  15. Types of Harness DIRECTIO N DO MAIN NATURE Feedforward Guides

    BEH AVI O UR I N F E RE N TI AL Feedback Sensors M AI NTAIN ABI L ITY ARC H I TECTURE DETERMINISTIC
  16. Types of Harness DIRECTIO N DO MAIN NATURE Feedforward Guides

    BEH AVI O UR I N F E RE N TI AL Feedback Sensors M AI NTAIN ABI L ITY ARC H I TECTURE DETERMINISTIC
  17. Types of Harness DIRECTIO N DO MAIN NATURE Feedforward Guides

    BEH AVI O UR I N F E RE N TI AL Feedback Sensors M AI NTAIN ABI L ITY ARC H I TECTURE DETERMINISTIC
  18. Types of Harness DIRECTIO N DO MAIN NATURE Feedforward Guides

    BEH AVI O UR I N F E RE N TI AL Feedback Sensors M AI NTAIN ABI L ITY ARC H I TECTURE DETERMINISTIC
  19. Types of Harness DIRECTIO N DO MAIN NATURE Feedforward Guides

    BEH AVI O UR I N F E RE N TI AL Feedback Sensors M AI NTAIN ABI L ITY ARC H I TECTURE DETERMINISTIC
  20. PRO M PTS AGEN TS.M D SPEC S, PRD &

    ADR STYL EGUI DES REF EREN C E DO CS RUL ES SC RI PTS / CL I TO O LS CO DEM O D S L AN GUAGE SERVERS ... Feedforward Guides I NI TIAL G E N E RATI O N  ✦ HUMAN AGEN T UNIT TESTS E2E TESTS STATI C AN ALYSI S REVI EW AGEN TS LO G S BROWSER L I N TERS SBO M VAL I DATI O N SEC URI TY SCAN N ERS ... Feedback Sensors SEL F - C O RRECTI N G
  21. Implementing Guides ✦ ✦ . . . . . .

    ✦ Write actual documentation ✦ Guides, rules, conventions; plus examples ✦ Current architecture overview ✦ Long-term specs, PRDs, and ADRs Add helpful tooling ✦ Code generation scripts, tools, helpers ✦ Language servers Entrypoint is the "router" my_app ├── AGENTS.md ├── docs/ │ ├── rules/ │ ├── guides/ │ ├── adrs/ │ └── specs/ ├── └──
  22. Implementing Sensors ✦ More important than Guides ✦ ✦ For

    maintainability and architectural quality Focus on Deterministic controls rst ✦ I M PLE M ENTATIO N L AYE RS 1. L I N TI N G & STATI C C H EC KS 2. UNIT TESTS Fast, reliable, cheap 3. I NTE G RATI O N/ E2 E Implementation Layers 4. AI REVI EWS ✦ Fastest & accurate feedback early ✦ Goal: Push agents' reliable coverage as far up as possible fi ✦ 5. MANUAL Q A
  23. Recommendations ✦ ✦ ✦ Establish Discipline ✦ Capture standard conventions,

    security mandates, architecture patterns ✦ Keep AI out of writing tests, preserve double-bookkeeping Build Reusable Harnesses ✦ CI templates with common deterministic checks ✦ Inferential review agents for security, architecture, gap analysis, even PR reviews Scale via Service Templates ✦ Service-level AGENTS.md
  24. Service Templates ✦ Enterprises & agencies have pre-de ned service

    templates ✦ Internal team guides ✦ Codemods & internal tools ✦ Boilerplate projects Embed harnesses directly in them ✦ Scaffold not just code, but AI knowledge and conventions from day one ✦ Inter-organization review agents fi ✦
  25. Advanced Workflows ✦ Custom skills and slash commands ✦ Subagents

    for sub-tasks for context optimization ✦ Agent Councils & Consensus ✦ ✦ Parallel agent execution with git worktrees ✦ ✦ Adverserial reviews with multiple agents deciding on next steps Multiply output using same harness Independently running agent loops ✦ Spec → Code → PR → Review → Address Feedback → Merge

More Decks by GDG Lahore

Other Decks in Programming