Agent Starter Pack: Google's Scaffolding for ADK Agents, Now in Maintenance Mode
Ship AI Agents to Google Cloud in minutes, not months. Production-ready templates with built-in CI/CD, evaluation, and observability.
At a glance
- What is it?
- The Agent Starter Pack generates a deployable Google Cloud agent project from six templates, wiring in Terraform, CI/CD, evaluation and observability. Its own README now points new projects elsewhere, which changes who should use it.
- Who is it for?
- Adopt Agent Starter Pack only if you are maintaining an existing project it already generated, or if you specifically need the adk_java or langgraph templates and accept that no new ones are coming. New projects should run uvx google-agents-cli setup instead, as the README directs.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 56 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap Agent Starter Pack fills between an ADK prototype and a deployed service
Writing a ReAct loop with Google's Agent Development Kit is a short afternoon. Getting that loop behind a load balancer, with a container image, a service account, a CI pipeline and trace export, is the part that eats the following two weeks. Agent Starter Pack exists for that second phase. It is a Python package that emits a complete project tree: backend agent code, a frontend, deployment infrastructure, and the configuration that ties them together. The README frames the split plainly: focus on your agent logic, and the pack supplies infrastructure, CI/CD, observability and security. The intended user is a Python developer who has already decided to run agents on Google Cloud and does not want to hand-assemble the surrounding plumbing. It is not aimed at someone still choosing a model provider, and it is not a runtime library you import into an existing service. The unit of delivery is a generated repository, not a dependency.
What the six templates actually contain
The template list is the concrete part of the offer. adk is a base ReAct agent on the Agent Development Kit. adk_a2a adds Agent2Agent protocol support for communication between distributed agents. agentic_rag does document retrieval and question answering over either Vertex AI Search or Vector Search, which is a meaningful fork: the two retrieval backends have different setup and cost profiles, and the template picks one at generation time. langgraph is the same ReAct shape built on LangChain's graph library rather than ADK, which matters if your team already has LangGraph code. adk_java targets the Java ADK, so the pack is not purely a Python story despite the repository's primary language. adk_live is a real-time multimodal agent handling audio, video and text through Gemini. The spread is deliberate: two orchestration frameworks, two retrieval backends, two languages, one streaming case. What is absent is equally informative. There is no template for a batch or offline agent, nothing for a non-Google model provider, and no minimal single-file example for someone who wants to read the smallest possible working agent before committing to a structure.
create and enhance: two entry points into the same scaffold
Getting started is one command. With uv installed, the README gives uvx agent-starter-pack create, which runs the generator without a permanent install. The pip path is the conventional one: python -m venv .venv, source .venv/bin/activate, pip install --upgrade agent-starter-pack, then agent-starter-pack create. The second entry point is for code you already have. Running uvx agent-starter-pack enhance from a project root adds the deployment and infrastructure layer to an existing agent instead of generating a new one. That distinction is the most useful thing in the README, because it means the pack is not only a greenfield tool. The README also points at Firebase Studio and Cloud Shell for zero-setup trials, and the repository layout shows the templates living under agent_starter_pack/agents/ with an idx resource directory used by the Firebase Studio template link. The documentation site covers installation, observability and deployment in separate guides, so the CLI is the entry point and the guides are where the deployment targets are explained.
The maintenance-mode notice is the first thing to read
The README opens with a warning banner rather than a feature list. Agent Starter Pack is in maintenance mode. Active development has moved to agents-cli, described there as the next evolution. The stated terms are specific: critical fixes only, no new features, no new templates, no new deployment targets. The README instructs new projects to start with agents-cli via uvx google-agents-cli setup and existing projects to migrate, and claims migration takes minutes with agent code, tests, Terraform and CI/CD carrying over without rewrites. That claim is the one worth testing yourself before trusting it, because it is a vendor statement about a different repository. The release cadence in the repository metadata is consistent with the banner: v0.41.3 landed in April 2026, with v0.41.1 and v0.41.2 in the same month, and the last push to the default branch is later than that. A cluster of patch releases followed by a maintenance declaration is what a wind-down looks like. For an engineer choosing a scaffold today, this reframes the whole decision: you are not picking the actively developed Google option, you are picking the one that has stopped growing.
Where the generated project stops being your friend
The failure mode of any scaffolder is the moment you need something the template did not anticipate. Here the constraint is explicit rather than hidden: no new templates and no new deployment targets are coming. If your agent needs a deployment shape outside Cloud Run and Agent Engine, the pack will not grow one for you. The second limitation is coupling. The generated tree includes Terraform and CI/CD configuration that you then own, and the README's own migration claim treats that Terraform as portable to agents-cli, which implies the two tools share assumptions about how infrastructure is described. If you fork the generated Terraform heavily, that portability is exactly what you erode. The third is the framework fork. Choosing langgraph or adk_java means choosing a template whose upstream framework moves independently of this repository, and with the pack frozen, framework updates land on you. None of this makes the output bad. It makes the output a starting point you must be willing to maintain without upstream help.
How it differs from assembling ADK or LangGraph yourself
The honest alternative is not a competing product. It is the ADK Samples repository, which the README links as a place to explore more ADK examples, plus your own Terraform. The difference is scope. ADK Samples show agent patterns; they do not generate a deployment pipeline, a monitoring configuration or an evaluation setup. Agent Starter Pack's value is that the boring layer arrives already wired, and the trade is that the wiring reflects someone else's defaults. A second alternative is agents-cli itself, which the README positions as the successor with a unified CLI replacing the Makefile, commands for run, deploy, eval run, eval compare, playground and lint, bundled coding-agent skills for Claude Code, Gemini CLI or Codex, and first-class support for Google Cloud's Agent Platform. Choosing between the two is not a feature comparison right now. It is a choice between the tool with a migration guide pointing away from it and the tool that guide points to. The only reason to stay is if a template here, such as adk_java or langgraph, has no equivalent there and you have verified that.
Licence, upgrade cost and what to check before you generate anything
The repository is Apache-2.0, which permits commercial use, modification and redistribution with the usual conditions around notices and patent grant. That covers the generator and the templates. It does not automatically cover every dependency a template pulls in, and the generated project will include libraries from Google, LangChain and others under their own terms, so read the generated dependency manifests rather than assuming the top-level licence settles everything. This is not legal advice. On upgrade cost, the picture is clear from the banner: critical fixes only, so security patches should still arrive while feature work will not. Budget for the generated Terraform and CI configuration as code you own outright. Before running create, decide which template matches your framework, because switching orchestration frameworks later means regenerating rather than editing. Then run the command in a scratch directory, read the Terraform and the pipeline definition it produces, and check that the deployment target you need is among the ones the pack supports. If it is not, the banner has already told you it never will be.
Editorial conclusion
Adopt Agent Starter Pack only if you are maintaining an existing project it already generated, or if you specifically need the adk_java or langgraph templates and accept that no new ones are coming. New projects should run uvx google-agents-cli setup instead, as the README directs. Before committing either way, run the create command in a scratch directory and read the generated Terraform and cloudbuild files, since those are what you will own after the scaffold is written.
Community notes