Gemini on Agent Platform vs Agent Runtime demo (Go)

[pollenjp]

Summary

Adds gemini-platform-vs-runtime-2026-05-14/ with two minimal Go clients and a mock server illustrating the difference between the two surfaces of the rebranded Gemini Enterprise Agent Platform (April 22, 2026 Cloud Next '26 rebrand of Vertex AI):

• Gemini on Agent Platform — stateless model API (POST /v1/models/{model}:generateContent). Client owns history, memory, tool execution. Shown both correctly (with client-side history) and incorrectly (without) to make the statelessness obvious.
• Agent Runtime (formerly Agent Engine) — managed runtime hosting an ADK agent. Client creates a session and sends one message per turn; the server keeps history, runs tools, and persists Memory Bank state.

Both clients target the same toy task (track running expense totals across 3 turns) so the diff is purely the architectural surface, not the task. A mockserver binary fakes both endpoints deterministically so the demo runs offline with no API key.

Layout

gemini-platform-vs-runtime-2026-05-14/
├── cmd/{mockserver,platform,runtime}/main.go
├── internal/api/api.go    # shared JSON shapes
└── README.md              # Japanese walkthrough + sources

Test plan

• go build ./... clean
• go vet ./... clean
• go run ./cmd/mockserver + go run ./cmd/platform — running totals work with history, break without
• go run ./cmd/runtime — session-based, server keeps state, no history sent by client

https://claude.ai/code/session_01CaNtg3DE1pcUwHLF7HErZX

---

Generated by Claude Code