Tracing a Google ADK multi-agent app with Opik: Flash-Lite 5x cheaper, 6x faster
dl_weekly · x · 2026-11-10
A DL Weekly tutorial shows how to trace and evaluate a Google ADK multi-agent support app with Opik, Comet's open-source LLM observability platform.
- Architecture: a supportagent (order lookup, help-center search, human escalation) with a refundsagent sub-agent (30-day policy check, refund execution) for a fictional outdoor gear store; one trace shows the full handoff plus every Gemini and tool call
- Key recipe: create an OpikTracer and call trackadkagentrecursive(rootagent, tracer); evaluation runs ADK Runner cases through opik.evaluation.evaluate, scoring both what tools the agent called and what it said
- Benchmark on 23 hard support cases: Gemini 3.5 Flash-Lite ran 5x cheaper and 6x faster than Gemini 3.8 Flash with equally safe refund behavior; an LLM-as-a-judge metric scored one correct refusal only 0.50
- Everything ships in a single repo that works from a fresh clone
More from coding & agent
- Stored prompt injection: payloads hidden in agent memory fire later in clean sessions — Future_AGI · 2026-11-10
- Visual Token Routing cuts computer-use agent inference cost up to 3.3x with 2.5-5.8x fewer tokens — wellecks · 2026-11-10
- Claude Fable 5.2 reportedly coordinated 1,000+ agents to build a bootable OS in 8 days — Spirited-Sir-3034 · 2026-11-10
- Ex-Zalando AI lead's thread: agentic AI changes what a company is, not just add-ons — ilparone · 2026-11-10
- PSAISuite Gives Claude Fable-5.2 Real PowerShell Hands for Agent Loops — dfinke · 2026-11-10
- Coding Assistants Just Moved Your Bottleneck: 27 PRs Stuck Waiting on Review — Pavan_Belagatti · 2026-11-10