DevFest Trabzon 2026 · Workshop guide · about 15 minutes to read
Building Trabzonspor's AI Analysis Room
The written version of the 90-minute hands-on workshop by Omar Elcircevi
Agent recipes with Google's Agent Development Kit. Build a football analysis staff out of one agent, a sequential team, a parallel team, a review loop and an orchestrator, all working on real Trabzonspor data.
It's Saturday night. Trabzonspor host Gaziantep FK tomorrow, Sunday 1 November 2026, 16:00 Istanbul time, in Round 10 of the Süper Lig. You're the club's new analysis staff, and the coaches want a briefing by morning: who the opponent is, how we play them, who starts, and what could go wrong.
You're going to build that staff out of agents. Each step adds one staff member, and each one is an agent pattern you'll keep using long after this match:
- The Scout: one agent with tools.
- Match Prep: agents that run in a fixed order.
- Scouting Department: agents that run at the same time.
- Tribün: two agents in a loop, one proposing, one criticising.
- Technical Director: an agent that delegates to all the others.
It all runs on Google's Agent Development Kit (ADK) for Python, a free Gemini API key, and real data from both teams' last ten matches. I ran this as a workshop at DevFest Trabzon 2026; this guide is for everyone who wasn't in the room.
The pages
- START · 1 MINSetupInstall uv, get a free Gemini API key, clone the repo and open the ADK dev UI.
- START · 1 MINThe dataThe FotMob and ESPN snapshot behind every agent, and why the tools return computed numbers.
- STEP 1 · 2 MINThe ScoutOne LlmAgent with five tools: an instruction, docstrings, and tools that do the counting.
- STEP 2 · 2 MINMatch PrepA SequentialAgent: three agents in a fixed order, handing work over with output_key and placeholders.
- STEP 3 · 2 MINScouting DepartmentA ParallelAgent inside a SequentialAgent: three scouts at once, then a writer merges their reports.
- STEP 4 · 2 MINTribünA LoopAgent: the lineup picker proposes, the home crowd objects, and exit_loop ends it.
- STEP 5 · 3 MINTechnical DirectorAn orchestrator with AgentTool: descriptions drive routing, and every agent has one parent.
- BONUS · 1 MINLive data over MCPThe same scout on a live football MCP server over SSE, with no change to the agent.
- WRAP-UP · 1 MINWhich recipe whenWhich agent pattern to use when, and what to try next.