Google Agent Development Kit (ADK)

You will like ADK if your shop is polyglot: it is the one framework on this board with Python, TypeScript, Go, Java and Kotlin, so the Java team and the Python team stop arguing and share one agent vocabulary. The daily trait is the dev UI, where you watch a multi-agent run step by step instead of reading logs.

The pull toward Google Cloud is constant, in the defaults, the deploy commands and the examples. Pick it if you are already there. Pick the OpenAI Agents SDK if you live on OpenAI's platform and want the same shape with fewer languages.

The risk is gravity. Gemini is the default, the documented MCP sample sets model="gemini-flash-latest", and everything else arrives through a LiteLLM adapter. Portable in principle, Google in every example, which means the first time a non-Gemini model misbehaves you are debugging an adapter nobody at Google runs in production.

The consequence for a team is that model portability is a claim to test on day one, not a feature to assume. Non-Gemini examples in the official docs would change this verdict. What it does right: an evaluation framework with custom metrics and user simulation ships with the toolkit, which most frameworks leave to the user.

The composition model is documented. 1. LlmAgent for model-driven steps. 2. Sequential, Parallel and Loop agents for deterministic control flow. 3. Graph-based workflows for explicit execution paths. 4. Context through structured sessions and memory with automatic token optimization, the vendor's phrase for a compaction policy the docs do not specify.

No benchmark is claimed. The consequence is that the deterministic scaffolding is auditable and the compaction is not, so a run's reproducibility depends on the part least described. The observation: making the deterministic agents first-class is the correct response to model nondeterminism.

Give away the SDK, own the runtime. One-command deployment to Agent Runtime, Cloud Run or GKE is where the meter starts, and Agent Runtime in particular is a managed service only Google sells, so every agent built here is a future line on a Google Cloud invoice. Moat: distribution through cloud accounts that already exist.

Pivot risk: Google consolidates agent surfaces often, and a framework can be folded into a platform SDK with a rename and a migration guide. The library will survive that; the deployment path may not. Position: long the cloud, hedge the API by keeping the agent definitions portable.

The demo is a multi-agent team traced in the adk web UI. As a dependency: maintained by Google, $0 for sixty engineers, and when agents run on Google's runtime, support and SLAs come through the cloud agreement procurement already signed, which turns a framework question into a line item we already have.

CI fit is headless by nature, a framework runs wherever Python or Go runs. Onboarding is a scaffolding CLI and a visual debugger, so a mid-level engineer has a working agent on day one. Approved with conditions: cloud terms reviewed for agent data, and the deploy target agreed before the first service ships.

Apache-2.0 across five languages, and go get google.golang.org/adk/v2 means I do not have to write Python to use it. McpToolset with StdioConnectionParams attaches any MCP server I already run, and it is one of the few frameworks that also exposes its own tools as an MCP server, so my ADK agents become tools for my other agents.

Ollama runs through LiteLLM, one adapter away, so the whole thing works on my box. A fork would be large but possible, and the five-language spread means a fork would need five teams. I would rather contribute upstream than own that.