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

[AgentConHelsinki2025]From Assistant to Agent: ...

[AgentConHelsinki2025]From Assistant to Agent: How Developer Agents Reshape Mobile Workflows

This is the slide, "From Assistant to Agent: How Developer Agents Reshape Mobile Workflows" at AgentCon Helsinki 2025.
#AgentConHelsinki hashtag#AgentCon hashtag#GenAI hashtag#aicommunity hashtag#techevent

Avatar for Kenichi Kambara

Kenichi Kambara

October 01, 2025
Tweet

More Decks by Kenichi Kambara

Other Decks in Technology

Transcript

  1. Kenichi Kambara (@korodroid) October 1, 2025 From Assistant to Agent:

    How Developer Agents Reshape Mobile Work fl ows AgentCon Helsinki 2025
  2. •Mobile App Development •Talks (22 Int’l/100+ Domestic) •Writings (10 Dev/Edu

    Books) •[Of fi cial] Evangelist at NTT TechnoCross e.g. AI-powered call analysis, livestock monitoring, and privileged access management •[Private] iplatform.org About me Kenichi Kambara (X:@korodroid)
  3. Agenda •AI Paradigm Shift •The Complexity of Mobile Apps Development

    •“Chat-based AI” vs. “Agent-based AI” •Use Cases (with Demo) •Practical Insights
  4. Evolving the Developer Work fl ow  $IBUCBTFE"* 🧠Gemini 🧠ChatGPT

    🧠Claude "HFOUCBTFE"* 🛠Cursor 🛠GitHub Copilot (agent-like) 🛠Claude Code Conversational interface for Q&A, guidance, and ideation. Autonomous, tool-using agents that plan, act, and deliver outcomes.
  5. AI Paradigm Shift  From Passive to Active Key Transitions

    Evolved from waiting for prompts to acting on context •From isolated commands to continuous collaboration •From single- fi le focus to project-wide understanding •From reactive suggestions to proactive improvements
  6. The Complexity of Mobile Apps Development  • Platform Support

    iOS Android Cross-platform Device diversity Coding Debugging Testing Refactoring Performance Versioning Review guidelines Metadata Release Accessibility Responsive design Localization Usability • Development Process • App Store Requirements • UI/UX
  7. Real Pains in Mobile Apps Development  🔁 Repetitive &

    cross-cutting changes ⚡ Slow iteration loops 🔎 Late discovery of issues 🧪 Thin tests 📱 Device/OS fragmentation
  8. “Chat-based AI” vs. “Agent-based AI”  Aspect Chat ‑ based

    AI Agent ‑ based AI Learning curve Easy to get started; simple chat interface. Higher learning effort; know its abilities and limits. Interaction model Conversation ‑ driven; explicit prompts; manual copy between tools Goal ‑ driven; plan → act → review; orchestrates tools/APIs Project awareness Limited context window; no project ‑ wide understanding Understands codebase & dependencies; workspace indexing Execution capability Suggests code; user applies changes; no direct code manipulation Direct edits, runs tests/builds; multi ‑ step operations across files Context & state Session resets disrupt continuity; weak persistence/ memory Persistent context across sessions; memory & artifacts Delivery workflow Human handoffs; PR/tests/reports done manually Draft PRs, tests, profiling & reports automated; human approval gate
  9. Deep Dive into “Agent-based AI”  Plan → Act →

    Check — context-aware, tool-using, outcome-driven Key Characteristics Context Awareness Multi ‑ step Reasoning Direct Execution Persistent State Understands codebase, dependencies, and developer intent. Plans tasks and executes across fi les/functions. Edits code, runs tests/builds, generates draft PRs. Keeps memory and artifacts across sessions.
  10. { "greeting": "Hello" } Create some new fi les lib/l10n/app_en.arb

    { "greeting": "͜Μʹͪ͸" } { "greeting": "ﺎً ﺒﺣﺮﻣ" } lib/l10n/app_ja.arb lib/l10n/app_ar.arb > fl utter pub get Automatically Generated…  Internationalization Flow (2/3) 1SFQBSF-BOHVBHFSFTPVSDFT
  11. <project>/lib/main.dart Text(AppLocalizations.of(context)!.greeting); import 'package:flutter_localizations/flutter_localizations.dart';  Internationalization Flow (3/3) $IBOHFUIFDPEF return

    MaterialApp( title: ‘App', localizationsDelegates: const [ GlobalMaterialLocalizations.delegate, GlobalWidgetsLocalizations.delegate, GlobalCupertinoLocalizations.delegate, ], supportedLocales: const [ Locale('en'), // English(LTR) Locale('ja'), // Japanese(LTR) Locale('ar'), // Arabic(RTL) ],
  12. Processing with Claude Code  You are an autonomous code

    agent in a Flutter repo. Goal: A screen scrolls slowly. Diagnose → improve → verify → open a draft PR. Proceed without approval. Guardrails: - Keep business logic & public APIs unchanged. - Limit edits to the affected screen (e.g., MyHomePage) and minimal helpers. - Keep the diff small and self-explanatory. Plan: - List 3 hypotheses for the slowdown. - Add lightweight measurements ( fi rst-frame builds, per-screen scroll rebuilds) using debug-only logs. - Start measuring immediately. Diagnose: - Run measurements ( fl utter run/test/analyze). - Output key numbers on one line and choose up to 2 approaches with a brief rationale. Act: - Apply a minimal design that reduces fi rst-frame builds and scroll rebuilds while preserving visuals/API. - Show the diff and apply it. Check: - Run analyze/test; report green/red. - Show Before → After numbers on one line each. Draft PR: - Branch: perf/scroll-optimization - Commits: perf(ui)… ; chore(debug) remove counters - Create a draft PR with problem, root cause (numbers), changes, diff size, key metrics, and checks. Start now. If anything fails, propose a fallback and continue. Plan → Diagnose → Act → Check → Draft PR
  13. Safe Agent-Based AI  Permission Scope Least privilege by default;

    disable risky actions. Data & secrets Use secret managers; hide personal data; mask tokens in logs. Prompt/tool safety Approved tools only; human approval for risky tasks. Quality & change control CI checks; audit logs; staged rollout. Cost & Ops Monitoring Monitor usage dashboards; alert on spikes and runaway tasks. Execution guardrails Set time/cost caps; enforce rate limits; auto-abort on failure.
  14. Evaluating Agent-based AI  Context Understanding Understands the project, sees

    how fi les connect, and gets the developer’s intent. Autonomy Level Works on its own with light guidance and makes safe choices within set rules. Multi-step Processing Plans and carries out several steps across fi les and functions to fi nish a task. Iterative Improvement Learns from feedback and improves solutions iteratively.
  15. The Evolution of Release Automation  #FGPSF $IBUCBTFE "GUFS "HFOUCBTFE

    “Instruct → execute → con fi rm” remains human-driven: answers, not completion. Release cycles depend on human availability. Give a goal → Agent plans → implements → tests → measures → creates a draft PR in an automated loop. Humans focus on approval. 💬 Chat-based AI 🧑💻 Human Work ⚙ CI ✅ Human Review 📝 Draft PR ✅ Approval Gate Plan Act Test Measure Loop
  16. Key Takeaways: 3 Keys to Agent-based AI •Read diffs as

    data (Small PRs) •Gate changes via draft PRs (Draft- fi rst) •Keep humans on risky edits (Risk-based reviews)