Agent Development Kit
Confidence 0.75 · 5 sources · last confirmed 2026-09-04
Google’s framework for building and orchestrating LLM agents (usually “ADK”), and the product that sits underneath most of the wiki’s Google-sourced agent material. It occupies the framework slot in the vocabulary Harrison Chase sharpened on the agent development lifecycle — frameworks / runtimes / harnesses / no-code — supplying agent composition primitives rather than a finished product surface. That distinguishes it from Antigravity, Google’s harness and IDE ecosystem, which the wiki tracks separately and which should not be conflated with ADK.
What the corpus shows it doing
- Wiring up the reason-act loop in practice. 2026-06-10-google-cloud-tech-ai-agents-explained-first-agent builds a self-correcting multi-agent blog-writing system in Python on ADK, pairing the ReAct theory with a hands-on build — the operational bookend to the 2022 ReAct paper. Three agent patterns are demonstrated: sequential, reactive, and planning.
- Composing loops into graphs. 2026-08-25-thurium-wang-google-cloud-four-ways-loop-engineering-fails is the ADK team’s account of when a single loop stops coping and the work should move to graph engineering — “you can put loop as part of the graph, or you can create your own nodes and edges.”
- Shared state as the graph’s traversal mechanism. 2026-09-03-thurium-wang-google-cloud-graph-engineering-101 names what ADK actually supplies for graph engineering: “if you build a multi-agent system we have this shared state among different agents; if you build a workflow, we have shared state among nodes in this workflow as well.” Two granularities, one primitive — and it is what makes “passing memory or information down the graph to the next node” work. The video ships with a companion codelab (
g.dev/ai/adk-2-orchestration) and the fan-out / join / router pattern vocabulary the corpus now uses for ADK workflows. - Alongside the CLI surface. 2026-04-22-cheung-ippolito-secchi-google-agents-cli covers Google’s agents CLI in the same product family.
Why it matters to the wiki
ADK is one of the few named frameworks the corpus can watch across a full arc — theory (ReAct), a build, a failure taxonomy, and an escalation path from loops to graphs — all from the same vendor within five months of 2026. That makes it a useful worked instance of agent-harness claims that are otherwise stated abstractly, with the standard caveat that every one of those sources is vendor-produced and none of them measures anything.