A Google for Developers blog post that draws five architectural patterns out of the Google Cloud AI Agent Bake-Off, a competition series in which developer teams were given tight time limits to build autonomous agents for real industry problems — e-commerce returns, legacy banking modernization, and go-to-market automation. The authors present the series as evidence that the practice has moved past prompting a single model: in their framing, getting from a demo to a production application is a matter of multi-agent architecture, state management, and deterministic guardrails rather than better prompt engineering.
The five patterns are offered as an architectural blueprint rather than as findings: decompose into specialized sub-agents behind a supervisor, expect the agent harness to be short-lived, treat multimodality as native, adopt open agent protocols instead of bespoke integration code, and confine the model to reasoning while deterministic code performs the work. Each is illustrated with an episode from the series, and the post closes by pointing readers at Google’s own Agent Development Kit (ADK) and its Agent Starter Pack.
Key Points
- Decompose a complex agent into specialized sub-agents with tightly scoped prompts, coordinated by a supervisor agent that routes traffic — the post treats agents as microservices and argues that a single large prompt handling intent extraction, retrieval and reasoning at once invites hallucination and latency spikes.
- The post reports that in Episode 3 a team running tightly-scoped agents in parallel cut their processing times “from 1 hour down to just 10 minutes”. This is a single result observed during the series, reported by the organizers, not a measured benchmark.
- Modular decomposition is also presented as a maintenance argument: swapping a model or changing a database schema touches one sub-agent rather than the whole workflow.
- Architect for impermanence — the authors argue that a complex agent harness written today may be superseded within months by model capability, citing their own e-commerce challenge, where a multi-step virtual try-on was built under a three-hour deadline. The post reports that the first version of Nano Banana was “just weeks away” at the time, and that today the same experience “can be achieved in a single prompt”.
- Treat multimodality as a core requirement rather than an add-on: the post argues that a text-only recommendation fails real retail problems, and that the stronger architectures ingested user photos, extracted visual context and triggered image-generation tools to compose a visual result.
- Adopt open agent protocols rather than writing custom API wrappers per tool. The post names MCP, A2A, UCP, AP2, A2UI and AG-UI as the current landscape and argues that mastering it is what separates prototypes from production systems.
- Reserve the LLM for reasoning and intent extraction, validate its output against a rigid schema (the post names Pydantic as an example), then hand the validated variables to deterministic code — a Python function or SQL query — to execute. The authors report that teams in the legacy banking challenge who let the model do financial arithmetic directly triggered validation errors.
Context
The post is written by Google staff, published on Google’s own developer blog, and its recommendations point at Google’s own frameworks, so its protocol and tooling advice is a vendor’s account of its own ecosystem rather than an independent comparison. Its evidence is a competition series the same organization ran: the episodes are described as firsthand observation of teams working under artificial time pressure, which the authors themselves frame as “raw and unfiltered” rather than as a controlled study.
The post presents its five patterns as a blueprint distilled from the series rather than as its own inventions, and does not claim to originate any of them.
Related terms in this wiki: Agentic Engineering, Manager Pattern, Sub-agent architecture, Guardrails.