From Multi-Agent Patterns
to Reliable Orchestration
Orchestration depends on one repeated decision: who should handle this piece of work?
Daniel Homola·Lead AI Engineer, BMW Research
My Recent Work on Agents at BMW
Recent focus
Agentic AI architecture
GUI agents that operate screens
Related talk
LLM-Based GUI Agents
AI Engineer Paris 2025
A Multi-Agent In-Car Assistant
Hypothetical example: what agents could naturally exist in any modern vehicle?
🗺️ Navigation Agent
Routing & destinations
🚗 Car Control Agent
Windows, climate, seats
🖥️ GUI / Computer-Use Agent
Operates apps via screen
The orchestrator chooses who acts, coordinates execution, and may combine results.
The Patterns: How Agents Coordinate
Patterns can be mixed. The challenge is the runtime decision.
Handoff
Control transfers.
Agents as Tools
Delegate work, await result.
Routing
Dispatch each turn.
Runtime decision:
which agent or tool should handle this piece of work?
Delegation Is More Than Tool Selection
Same selection interface, different decision.
Tool Selection
"Which tool fits?"
Tools should not overlap
One label should work
Agent Delegation
"Who should handle this work?"
Overlap is normal
Cost, speed, reliability & UX matter
User request: "Play some jazz"
Empty screen → Media API. Jazz playlist visible → GUI tap can also be valid.
Overlap + context create alternatives.
Agents are not tools
Even when delegation is implemented as a tool call, it isn't one.
Tool Call
Agent Delegation
Same tool-calling interface. Different control semantics.
Where the Decision Gets Ambiguous
The GUI agent is a useful case study: it can reach outcomes through the same UI the user sees.
| User Request | GUI Path | Specialist Path | Best? |
|---|---|---|---|
| "Navigate to Munich" | Maps UI → go | Navigation API | Specialist ✓ |
| "Close the windows" | Settings UI | Vehicle API | Specialist ✓ |
| "Open that restaurant's site" | Use browser screen | No specialist for this | GUI Agent ✓ |
| "Play some jazz" (music app open) | Tap visible playlist | Media API | Ambiguous ⚖️ |
A Possible Way to Evaluate Orchestration & Delegation
A direction, not a recipe.
User request + full runtime context (history, app/vehicle state, available agents & permissions) → acceptable agent/tool choices.
Accept valid choices, but evaluate cost, speed, reliability, and UX.
Enterprise path: each subteam builds evals for the agent it owns.
Together with the orchestrator team, they add cross-domain collaboration cases; the union becomes one shared orchestration & delegation benchmark.
Reliable orchestration depends on
reliable delegation decisions.
Who should handle this piece of work?
Daniel Homola · BMW Research
Let's connect
