Beyond Vibe Coding: Architecting the Future of Mobile Apps with AI Agents

A three-layer framework for shipping production mobile apps with AI coding agents — Agent Setup, Governance & Verification, and Agent-Driven Planning.

Transcript

1. The Agentic Mobile Era
2. 2 developers 2 months 2 native apps 150+ screens
4. The beginning
5. It starts like this … • Invents APIs that don't exist • Imports packages you never installed • Rewrites three files to fix one bug
6. Zero Onboarding • No project context • No coding standards • No architecture • No self-verification
8. Three Layers AGENT-DRIVEN PLANNING 03 GOVERNANCE & VERIFICATION 02 AGENT SETUP 01
9. AGENT-DRIVEN PLANNING 03 GOVERNANCE & VERIFICATION 02 AGENT SETUP 01
10. Fix Instructions, not the output
11. The Dosing Problem Too little Too much Hallucinations Lost focus
12. Context • What you write: prompt, instructions • What's loaded: skills, MCPs • What's given: system prompt, environment
13. Progressive • Core rules, always loaded • Details on demand • Live data at runtime Context
14. Onboarding the Agent • Who the agent is • What the project looks like • What it must always do
15. The Toolkit Skills • Static instructions • Loaded as headers • Cheap on context MCP • Live data access • Queried at runtime • Expensive on context
16. Our Setup Agentic Instructions ├── AGENTS.md (foundation — <100 lines) │ └── Reference Docs (on demand, one-line triggers) │ ├── domain_entities.md │ ├── technical_architecture.md │ ├── coding_rules.md │ └── ... ├── skills/ (operations, workflow, code guides) ├── templates/ (PRD, techspec, task, task-list) └── .aiexclude (context boundary)
17. Inside AGENTS.md • Role definition • Core principles and rules • Reference docs (on demand) • Mandatory steps
18. The Skill Library Operations • Build • Test • Sca f fold modules
19. The Skill Library Operations • Build • Test • Sca f fold modules Guides • Swi f tUI Views • Database • Navigation
20. The Skill Library Operations • Build • Test • Sca f fold modules Guides • Swi f tUI Views • Database • Navigation Workflow • Create PR • PRD • Tech Spec
21. Lessons learned Context Overload • All docs loaded at once • Critical rules ignored • Less context, be t ter results
22. Lessons learned Know Your Edge Cases • KMP: agent edited wrong pla t form • Unique project quirks need explicit rules
24. AGENT-DRIVEN PLANNING 03 GOVERNANCE & VERIFICATION 02 AGENT SETUP 01
25. Diff looked fine
26. Fast and broken • Agents never slow down on their own • Wrong direction compounds fast • Verification is the trust layer
27. Closing the loop Write Build Test Format Validate
28. VERIFY AT THREE LEVELS 1. Build + tests a f ter every change 2. Lint and format, always 3. AI-powered PR review
29. Our Setup • Mandatory build + format a f ter every change • Snapshot tests against Figma designs • CI on every PR • CodeRabbit for AI review
30. Lessons learned code Looked Correct • Agent handed o f f broken code • Developer fixing what agent should catch • Small issues, constant interruptions
32. AGENT-DRIVEN PLANNING 03 GOVERNANCE & VERIFICATION 02 AGENT SETUP 01
33. DEFINE THE PROBLEM, NOT THE SOLUTION
34. How It Usually Goes 1. Developer reads the ticket 2. Developer writes the prompt 3. Gaps in the ticket stay gaps in the code
35. Requirements First • Agent asks what the ticket doesn't • Requirements refined before code • Fewer surprises, fewer rewrites
36. From Idea to Implementation 1. PRD 2. Technical specification 3. Task breakdown 4. Orchestrated implementation
37. PRD: WHAT, NOT HOW • User stories, not code • Scope and constraints definition • Agent asks questions /prd Agent asks Developer answers PRD document
38. Tech Spec • PRD meets your codebase • Concrete entities • Grounded in existing pa t terns PRD Codebase Tech Spec
39. Task breakdown • Tech spec broken into steps • Dependency order defined • Each task: dedicated sub-agent & session Tech Spec Task 1 Task 2 ...
40. Build in parallel • Agents implement tasks • Parallel when possible • Verified at every step Tasks Agent Agent Agent ✓ ✓ ✓ Verified Build
41. Lessons learned Skipped the Requirements • Jumped from ticket to code • Agent built what was asked, not what was needed
42. Let's See It
43. • Define what to build • Human decides what ma t ters • Agent takes it from here Step 1: The Ticket
44. • Agent reads AGENTS.md • Only relevant docs pulled in • Lean context, sharp focus Step 2: Context Loads
45. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
46. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
47. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
48. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
49. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
50. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
51. • Refine requirements • Spec meets codebase • Tasks scoped and ordered Step 3: Plan the Work
52. • Plan approved, code starts • Follows your architecture • Skills guide every step Step 4: Agent Executes
53. • Plan approved, code starts • Follows your architecture • Skills guide every step Step 4: Agent Executes
54. • Build passes before hando f f • Tests run automatically • Code is clean when you see it Step 5: Self-Validates
55. • Review the code di f f • Run on device or simulator Step 6: you Validate
56. • Review the code di f f • Run on device or simulator Step 6: you Validate
57. • Agent opens PR • Changelog is auto-generated • CodeRabbit reviews changes Step 7: Open the PR
58. • Code is tested & forma t ted • Changes are documented • Team reviews Step 8: PR Ready
59. 2 developers. 2 months. 2 native apps. 150+ screens.
60. AGENT-DRIVEN PLANNING 03 GOVERNANCE & VERIFICATION 02 AGENT SETUP 01
61. first thing Monday 1. Create an AGENTS.md 2. Add one rule: always run build 3. Next feature, define requirements first
62. Agents run fast. Your job is to point them the right way. One layer at a time. h t tps://tomasparizek.com/talks/aiheroes-2026/ h t tps://www.linkedin.com/in/tomas-parizek/ h t tps://www.strv.com/