NS Agentics Foundation Meetup #7

Описание: Serbian Agentics Foundation Meetup #7 | ADR Validation, Goal Planning & Implementation Swarms

🗓️ February 5, 2026 | StartIt AI Hub, Novi Sad, Serbia

In this seventh meetup of the Serbian Agentics Foundation chapter, we validated and improved our Architectural Decision Records (ADRs) from Meetup #6, created a goal-oriented implementation plan, and launched agent swarms to begin building the community social network using domain-driven design.

What we covered:

Validated ADRs using the Requirements Validator agent — scored 73/100, identifying missing concrete values, ambiguous password requirements, and incomplete event definitions
Launched a swarm of 9 parallel agents to fix ADR weaknesses, which also created 3 new ADRs for missing areas (front-end architecture, file storage, email delivery)
Manually reviewed and improved password requirements — increased minimum from 8 to 12 characters, advocating for passphrases over passwords
Discovered dead letter queue strategy and event schema versioning as new architectural patterns added by agents
Used Goal Planner agent to create a phased implementation plan: 7 TDD-bounded contexts, 10 phases, 45 parallel work streams, 8-10 concurrent agents, ~180 files estimated
Started implementation swarm with the guiding principle: "Implement, integrate, verify. Repeat until done. No shortcuts — we value quality we provide to our users."
Introduced the concept of a "North Star" for agents — ADRs provide the what and why, agents decide how to achieve the goals
Demonstrated Cloud Code Web sandbox setup — preparing environments that persist between sessions, accessible from mobile
Installed compahook — a community-built tool from Agentics Foundation for better conversation compacting in Claude Code
Showed custom status line tracking project-specific metrics: 12 quality domains, 50+ agents, learned patterns, ADRs, and test counts
Explained Tiny Dancer hybrid router — routes tasks by complexity to different models (Haiku for simple, Sonnet for medium, Opus for complex)
Discussed CloudFlow plugins system — Agentic QE Fleet and Gas Town are now installable as CloudFlow plugins
Demonstrated different swarm types: hierarchical, hierarchical mesh, mesh, and star topologies based on task type

Tools used:

Claude Code (CLI + Web sandbox) & Cloud Flow V3 (orchestration with roof swarm)
Agentic QE Fleet V3 (quality analysis with DDD architecture)
Requirements Validator agent & Goal Planner agent
CompaHook (better compacting for Claude Code conversations)
DevPod + Docker — secured VS Code Server environments
Whisper (voice-to-text for prompting)

Key learnings:

Always validate ADRs before starting implementation — our first pass scored 73/100 and needed significant improvements
Agent swarms can fix multiple ADR issues in parallel — 9 agents fixed existing ADRs and created 3 new ones
Goal Planner creates phased plans with dependency graphs, file ownership, and quality gates — essential before launching implementation swarms
The "North Star" approach: give agents architectural guidance (what & why via ADRs), let them decide how to implement
Custom status lines give real-time project awareness — track domains, agents, patterns, tests, and vector counts
Tiny Dancer hybrid router saves tokens by matching task complexity to model capability
Cloud Code Web sandbox persists your environment — prepare once, continue from anywhere, including mobile
Cloud Flow documentation is written for agents to read, not humans — ask Claude to explain settings rather than reading the docs yourself
Track everything in version control — monitor what agents change through git diffs and delta PRs

Q&A highlights:

How much does this cost? On max subscription, no API costs. Estimated API equivalent: ~$200 during the session, ~$1,000 for a full working day. Tiny Dancer hybrid router optimizes by sending simple tasks to cheaper models
Can you use local models? Yes — both Cloud Flow and Agentic QE Fleet support Ollama, VLLM, and Open Router alongside Anthropic models
Why can't I find CloudFlow documentation? It's written for agents, not humans. Ask Claude Code to configure settings or explain features rather than reading the docs directly
What happened with the swarm when CloudFlow MCP failed? It automatically fell back to ruvswarm MCP, and the Agentic QE Fleet's queen coordinator can also start swarms independently

Next: Continue implementation from Phase 1 scaffolding, with agents building out the monorepo structure guided by 16+ ADRs and the phased implementation plan

📺 Watch previous Agentics Foundation content: video.agentics.org

🔗 Connect:   / dragan-spiridonov  

🌐 Agentics Foundation: agentics.org

#AgenticAI #AgenticQE #AgenticsFoundation #OpenSource #Serbia #AI #SoftwareDevelopment #DomainDrivenDesign #ClaudeCode #ADR #TDD