System Architecture in the AI Era - Build it Club 2026

Generating code has never been easier, but engineering a production-ready system remains hard. Surviving the clash with real users, strict budgets, compliance regulations, security risks and performance requirements demands more than just effective prompting—especially when AI stops being just a tool in your IDE and becomes a component in your architecture.

Join this session for an architect’s deep dive into the dual challenge of the AI revolution: building with AI and building for AI.

We will discuss the importance of engineering best practices in an era where more code is generated for us, shifting our role towards "architects of intent”. We’ll tackle the rise of agentic systems, manage the friction where deterministic logic meets probabilistic models, and redefine classic architectural quality attributes for the modern, AI-infused stack. All of this will be grounded in real-world examples leveraging low-level technologies and frameworks within the JVM ecosystem.

Transcript

1. BUILD IT CLUB System Architecture in the AI Era COMSENITY
2. Dual challenge of AI revolution COMSENITY
3. Building with AI vs Building for AI COMSENITY
4. Who am I • Pragmatic Software Architect • Creating software for nearly 20 years • Founder of COMSENITY • Helping companies to deliver real value • Occasional speaker and trainer LET'S CONNECT @piotrl sk wiec @piotrl sk wiec.bsky.soci l a a a a a COMSENITY
5. Building with AI COMSENITY
6. Code is cheap. COMSENITY
7. Congratulations, you’re an architect now. COMSENITY
8. Architecture
9. „You can outsource your thinking but you can't outsource your understanding.” - Andrej Karpathy COMSENITY
10. AI
11. f AI = AmplI ier
12. Fuel for the machine • Architecture Decision Records (ADRs) • Di gr ms nd context • Speci ic tions • AGENTS.md • Never trust, lw ys verify • ArchUnit • Fitness functions a a a a a f a COMSENITY
13. Everything that doesn’t it into a context window is now priceless. f COMSENITY
14. Embrace review process. COMSENITY
15. Demo #1
16. Building for AI COMSENITY
17. Fr mework Integr tion Observ bility Test bility System Inter ctions Context D t Not st nd rdized Security … Prompt Roles User/System Ch t Memory a a a a a a a a a a a a a a a V rious Input Types REST API Request/Response Stre ming St teless Pre-tr ined Completion Structured Output V rious Output Types
18. LangChain4j
19. modelcontextprotocol.io
20. • Don ted to the Agentic AI Found tion (hosted by the Linux Found tion) • >10,000 public MCP servers s of September 2026 (~1200 in Q1 2025) • 41% of enterprise AI te ms report t le st one MCP-b cked gent in production • MCP Register, MCP Bundles, MCP Apps… • 3 m jor spec revisions since M y 2025 (June nd November 2025, July 2026) • Structure tool output • Elicit tion • OAuth overh ul (MCP servers re now OAuth Resource Servers) + OpenID Connect Discovery • T sks (still experiment l) a a a a a a a a a a a a a a a a a a a a a a Source of st ts: https://www.digit l pplied.com/blog/mcp- doption-st tistics-2026-model-context-protocol
21. • C p bilities • Resources - generic text/bin ry d t ( ile content, API responses, UI etc.) • Tools - execut ble function lities • Prompts - reus ble templ tes for users • Elicit tion (new) - servers c n sk the user for missing input mid-c ll f a a f a a a a a a a a f a a a a a a a a a a a a a a a a a a a a a • T sks (new, extension) - c ll now, fetch l ter - sync h ndles for long-running server work • S mpling (deprec ted) - lips the low MCP server sks the client's LLM for completion, with the hum n in the loop • St teless • Tr nsports • Stdio • Stre m ble HTTP • HTTP/SSE (deprec ted) • D t form t - JSON-RPC 2.0
22. Swimming in acronyms soup AP2 A2A MCP Apps (MCP-UI) MCP UPC A2UI ACP XD AI-xD AG-UI
23. AI meets UI MCP Apps (MCP-UI) A2UI AG-UI
24. Demo #2
26. Agentic AI
27. Agentic AI Let LLM do the job • Agentic AI softw re - pursue complex go ls with high level of utonomy • Orchestr tion str tegies: • Work low - LLMs nd Tools re progr mm tic lly orchestr ted (well-de ined control low) • P tterns: sequenti l, loop, p r llel, condition l • „Predict ble” output, e sier to test, its for well-de ined t sks, less lexible • Pure Agents - LLMs fully control the low • LLM is responsible for choosing how to solve given problem using v il ble tools f a f a a f a a a f a a a a a a a f f a a a a a a a a a a a a f a • Flexible, even less predict ble, h rder to control
29. CPU:GPU ratio 1:4-1:8
30. CPU:GPU ratio 1:1-1:2
31. Agentic AI in 2026 - Reality Check Hype is loud. Production is harder. • 72% of enterprises re lre dy using or testing AI gents • 17% of org niz tions h ve deployed AI gents to prod, 60%+ pl n to within two ye rs • 46% s y the #1 ch llenge isn't model intelligence - it's integr tion with existing systems • 38% still keep hum ns-in-the-loop s the def ult control p ttern a a a a a a a a a a a a a a a a a a a a a a a Source of st ts: G rtner Hype Cycle for Agentic AI 2026 · Z pier St te of Agentic AI Adoption 2026 · 2026 St te of AI Agents Report · Snyk 2026 St te of Agentic AI Adoption
32. Journey in the search of complexity a a a a Monoliths Too E sy Microservices Too H rd Modul r Monoliths Too E sy NonDeterministic Interconnected Agents
33. L ngCh in4j LLM-driven, Work low P tterns L ngGr ph4j Cyclic l St teful Gr phs Google ADK LLM-driven, Hier rchic l Work low/LLM Agents Emb bel Go l-Oriented Action Pl nning (GOAP) nd Observe-Orient-Decide-Act (OODA) Loop Koog LLM-driven, GOAP pl nning, Finite St te M chines (FSMs) a LLM-driven, Work low P tterns a Spring AI a f a a a a a a f f a a a a a a a a a Agentic Frameworks
34. Quality attributes • Security • Test bility • Correctness • Perform nce a a a • Observ bility
35. Security • Att cks • Prompt Injection • Prompt Le king • J ilbre king • DoS • Defense is in your h nds a a a a a a a • Gu rdr ils
36. CVEs CVE2026 -355 68 8 3 7 2 2 6 2 0 2 E V C 4 9 9 4 5 5 CVE-202 4 0 6 1 2 5 2 CVE-20 CV E-2 02 CVE2026 -2274 4 5-6 814 3 0 3 7 2 2 6 2 0 2 E CV 5 2 0 2 E CV 7 6 9 3 5 3 4 7 2 2 6 2 0 2 E V C 1 1 7 2 3 5 2 0 CVE-2 CVE-20 26-2274 2 4 1 5 6 5 2 0 CVE-2 Source: thephilosophersmeme.com
38. Lethal Trifecta Access to Sensitive Data Exposure to Untrusted Content Vulnerability Ability to Externally Communicate
39. Performance • Perform nce-Oriented Prompt Engineering (sic!) • SLAs in non-deterministic world m y be h rd to re ch a a a a a • Lot of moving p rts
40. • Simple checks • All required inform tion provided in the nswer • Proper tools were invoked • Token counts • How to test something th t is inherently not deterministic? • M ke it more deterministic a a a a a f f a • Fight ire with ire - AI Model Ev lu tion - LLMs- -Judge a a Testability
41. Correctness • H llucin tions • Writing good prompt is n rt • Provide qu lity context to the model • Retries • Rollb cks (Compens ting Actions) a a a a a a a a a • HITL (Hum n in the Loop)
42. Observability • Support for metrics nd tr ces (Sem ntic Tr ces) • Follows OTel „Sem ntic conventions for gener tive AI systems” (still in development) a a a a a a • Cost-control
44. COMSENITY Thank you for your attention. Contact piotr.laskawiec@comsenity.com Social Media @piotrlaskawiec @piotrlaskawiec.bsky.social LET'S CONNECT