All 48 articles, sorted alphabetically
Test Agent Model Call Limits Before Upgrading to ADK Go 2.7
ADK Go v2.7.0, released on 2 November 2026, caps model calls per invocation at 500, makes optional placeholders surface errors, gives each credential its own MCP session and switches telemetry to GenAI v2. What the source does and what to check before upgrading.
Read article →ADK Python 2.12.0: Confirmation Gates on Workflow Nodes, Tool-Output Screening With Model Armor, Runtime Service Accounts and a Faster Long Session
google-adk 2.12.0 shipped on 13 October 2026. What require_confirmation on @node and FunctionNode does, how the Model Armor plugin now screens tool output through after_tool_callback, the new runtime service account for Agent Engine deploys, how to read the 40 percent claim, and what to re-test.
Read article →ADK + A2A interop
How Google ADK interoperates with the A2A protocol: exposing an ADK agent tree as an A2A service, publishing an agent card and skills, consuming a rem…
Read article →Agent-as-tool -- composing agents as callable capabilities
Deep-dive on the agent-as-tool composition pattern: a parent calling a sub-agent wrapped as a tool, call-and-return vs sub-agent transfer, encapsulati…
Read article →ADK Agent Lifecycle
A timeline of one ADK agent invocation, end to end: the user message reaching the Runner, the InvocationContext, agent.run_async and the async event l…
Read article →ADK Agent Types
A decision-framework guide to the four families of Google ADK agent: the reasoning, tool-using LlmAgent; the deterministic Workflow agents (Sequential…
Read article →The ADK Artifact Service
Deep-dive on ADK's Artifact Service: named, versioned, immutable blobs stored out of band in a pluggable backend (in-memory or GCS), referenced f…
Read article →ADK caching
Caching for Google ADK agents at every layer: provider-side context and prompt caching and what actually makes a prefix cacheable, request assembly th…
Read article →ADK callbacks architecture
How ADK callbacks actually work: the six before/after hooks around agent, model, and tool; the return-None-or-replace contract; CallbackContext and To…
Read article →ADK circuit breaker architecture
Resilience patterns for ADK agents: the closed/open/half-open circuit breaker with concrete rate-and-volume thresholds, why naive retry amplifies an o…
Read article →ADK code executor -- letting the agent run code safely
Deep-dive on the ADK code executor: the computation need, the code execution tool (agent writes and runs code), the sandbox, result feedback, iteratin…
Read article →ADK Context Compaction: Architecture Deep-Dive
How ADK reconciles unbounded session history with a bounded context window — token budgeting, pinned invariants, the elision-to-summarization ladder, …
Read article →ADK deployment architecture
Getting a Google ADK agent to production: the agent package layout, containerizing the API server, the three deployment targets (Agent Engine, Cloud R…
Read article →Building an Agent Eval Harness
A hands-on reference for the ADK evaluation harness: the two artifacts (a lightweight .test.json vs a .evalset.json of many cases), what an eval case …
Read article →ADK evaluation architecture
The architecture of evaluating ADK agents: eval sets and replay, the two scoring axes of tool trajectory versus final response, designing a rubric an …
Read article →ADK Events
A deep dive on ADK Events: the Event object (author, content, actions, invocation_id, timestamp), how agents yield events as an async stream, how the …
Read article →ADK guardrail architecture
Policy design for ADK agent guardrails: the capability envelope, provenance and trust boundaries, input screening, tool-use authorization by subject a…
Read article →ADK Guardrails Pattern
Hands-on implementation patterns for agent guardrails in Google ADK using callbacks: validate and sanitize user input in before_model_callback and sho…
Read article →ADK LlmAgent
A deep dive on ADK's LlmAgent: the instruction/system prompt and how session state is templated into it, model selection across Gemini and model-…
Read article →ADK long-running tools
Deep-dive on long-running tools in an agent development kit: a tool that starts async work and returns an operation handle, session pause with durable…
Read article →ADK Memory
A developer's guide to memory in Google's ADK: the distinction between short-term memory (a single Session's events and state) and long…
Read article →ADK memory bank
How a long-term memory bank works behind ADK's MemoryService: where the Session/State boundary falls, what is worth persisting, extracting facts …
Read article →ADK model fallback architecture
Deep-dive on model fallback in the Agent Development Kit: the swappable model abstraction, a health check that classifies failures into retry/next-tie…
Read article →ADK multi-agent architecture
How to compose multiple agents in Google ADK: the sub_agents hierarchy and its one-parent rule, LLM transfer versus AgentTool, description-driven rout…
Read article →ADK multimodal
A practical guide to non-text content in Google ADK: how images, audio, video and documents are typed as content parts, inline blobs versus file URIs …
Read article →ADK observability architecture
Running Google ADK agents observably in production: the four-lens architecture (events, traces, metrics, quality signals), the correlation contract th…
Read article →ADK Observability and Traces
A deep dive on tracing Google ADK agents with OpenTelemetry: the span tree a single invocation produces (agent span, LLM call spans, tool call spans, …
Read article →Google ADK Overview
An orientation to Google's Agent Development Kit (ADK): what it is and why it exists, its code-first Python and Java model, the core building blo…
Read article →ADK planning architecture
Planning inside a single ADK agent: ReAct-style interleaved reasoning versus plan-then-execute, what the planner slot on an LlmAgent actually injects,…
Read article →ADK plugin system architecture
Deep-dive on the Agent Development Kit plugin system: registering plugins with the Runner to run before/after agent, tool, and model callbacks across …
Read article →ADK RAG grounding
Deep-dive on RAG grounding in Google ADK: retrieval as a tool versus retrieval built into the turn, chunking and indexing choices, hybrid search and r…
Read article →ADK runtime architecture
How the Google ADK runtime actually executes: the Runner as entry point, run_async as an async generator of events, invocation context, what an Event …
Read article →ADK Safety Evaluation
How to evaluate agent safety in Google's ADK: why safety eval differs from capability eval, building adversarial and red-team test sets (jailbrea…
Read article →Sandboxing ADK agents
Execution isolation for ADK agents: why model-generated code must never run in your process, the isolation ladder from exec() to containers to gVisor …
Read article →ADK session resumption architecture
Deep-dive on ADK session resumption: the session store and append-only event log, checkpointing on each event and recovering on reconnect, resume toke…
Read article →ADK Session Storage
A production guide to ADK session storage backends: InMemorySessionService for dev and tests, DatabaseSessionService on SQLAlchemy (Postgres/MySQL/SQL…
Read article →ADK Sessions
A deep dive on ADK Sessions: the Session object as the unit of a conversation, what it holds (id, app_name, user_id, the ordered events list, the stat…
Read article →ADK sessions, state, and memory
How Google ADK models a conversation: the Session as the unit of a thread, the append-only event list as the durable record, State as a fold of state …
Read article →ADK Skills
What a ‘skill’ means in Google ADK: not a formal class, but a reuse and composition pattern — bundling cohesive tools, a focused instruction, and opti…
Read article →ADK State Management
A deep dive on state mechanics in Google's Agent Development Kit: the session.state dict as a shared scratchpad, the session / user: / app: / tem…
Read article →ADK streaming architecture
How streaming actually works in ADK: the event stream as the API, partial versus aggregated events, StreamingMode SSE and BIDI, run_live and the LiveR…
Read article →Tool authentication architecture
Deep-dive on tool authentication in agent frameworks like ADK: how an agent calls authenticated external APIs and MCP servers without ever holding a r…
Read article →ADK Tool Design Patterns
Design patterns for building good ADK tools: the docstring and type hints as the contract the model reads, returning structured dicts with a status fi…
Read article →ADK tool architecture
How the ADK tool system actually works: what a tool is under the BaseTool abstraction, how function declarations are generated from type hints and doc…
Read article →Building a Voice Agent with ADK
A hands-on guide to building voice and live-audio agents with Google ADK: native audio versus cascaded ASR-LLM-TTS, the run_live bidirectional runtime…
Read article →ADK workflow agents architecture
Deep-dive on ADK workflow agents: SequentialAgent, ParallelAgent, and LoopAgent as deterministic control-flow primitives wrapping LLM sub-agents, sess…
Read article →Scaling ADK Agents on Vertex AI Agent Engine Runtime
A deep dive on scaling ADK agents on Google Cloud's Vertex AI Agent Engine: serverless vs orchestrated runtimes, cost optimization with Cloud Run…
Read article →State Management and Callbacks in Complex Asynchronous ADK Workflows
In a sophisticated agentic system, many valuable tasks are not instantaneous.
Read article →