Where would you
build this?

Long-running query jobs run asynchronously for up to seven days; StreamQuery requests are capped at 10 minutes. Mid-job instance loss is not described.

Request-based billing allocates CPU during request processing; instance-based billing allocates it for the instance lifecycle. Idle instances can be shut down at any time.

Request-based billing allocates CPU during request processing; instance-based billing allocates it for the instance lifecycle. Idle instances can be shut down at any time.

RequestInput pauses a node and the reply becomes its output. The page does not state how long a pause can last or its storage needs.

RequestInput pauses a node and the reply becomes its output. The page does not state how long a pause can last or its storage needs.

interrupt() saves state through a checkpointer and the graph waits indefinitely until resumed with Command(resume=...).

ADK provides resume by invocation id; Agent Runtime pages don't describe it. Detecting the failure and triggering the resume are not described.

ADK provides resume by invocation id with ResumabilityConfig. Detecting the failure and triggering the resume are not described.

Sync durability writes each checkpoint before the next step; async risks an unwritten checkpoint on a crash. Detecting a crash is not described.

ADK's resume page says a stopped workflow must not be modified before resuming; adding or removing agents is not supported. No version-migration guidance was found.

ADK's resume page says a stopped workflow must not be modified before resuming; adding or removing agents is not supported. No version-migration guidance was found.

Edge changes between kept nodes are safe. Renaming or removing a node a thread is paused at or about to enter breaks its resume.

ADK RequestInput pauses a workflow node for human input and passes the reply on. Approver identity and a recheck before the action are not described.

ADK RequestInput pauses a workflow node for human input and passes the reply on. Approver identity and a recheck before the action are not described.

Human-in-the-loop middleware pauses a tool call for approve, edit, reject or respond decisions. Approver identity and a recheck before the action are not described.

Agent identity is a per-agent, least-privilege identity with certificate-bound credentials. A separate write-credential service outside the agent is not described.

ADK tools act with the agent's own identity, such as a service account, or the controlling user's through OAuth. No separate write service is described.

The guardrails page describes no credential handling for tools and no separate write service.

ADK resume reinstates results of tools that returned and re-runs the failed one. Tools may run more than once; ADK says to prevent duplicate runs.

ADK resume reinstates results of tools that returned and re-runs the failed one. Tools may run more than once; ADK says to prevent duplicate runs.

Completed task results load from the checkpoint on resume. A task that started but did not finish may run again, so make side effects idempotent.

RunConfig max_llm_calls caps LLM calls per run (default 500). The RunConfig reference lists no token or cost limit.

RunConfig max_llm_calls caps LLM calls per run (default 500). The RunConfig reference lists no token or cost limit.

Model call limit middleware caps model calls per run or per thread, ending gracefully or raising an error. Its parameters are call counts.

Agent Platform Data Access audit logs include session creates and event appends but are disabled by default. Approver-level records were not found.

Cloud Run writes Admin Activity, Data Access and System Event audit logs for service and job operations. No documented per-run approval record was found.

Cloud Run writes Admin Activity, Data Access and System Event audit logs for service and job operations. No documented per-run approval record was found.

Agent Gateway applies Model Armor templates to agent prompts and responses and can redact or block content that violates them. Stored-data redaction is not described.

Model Armor plugin screens user input and model output for sensitive data and blocks flagged content, showing replacement text. Stored-data redaction is not described.

PIIMiddleware redacts, masks, hashes or blocks email, card, IP, MAC, URL or custom-regex matches in input, output or tool results. Storage paths are not described.

ADK supports Gemini, Claude, Agent Platform-hosted models, and LiteLLM, Ollama and vLLM connectors. Model routing includes automatic failover on error.

ADK supports Gemini, Claude, Agent Platform-hosted models, and LiteLLM, Ollama and vLLM connectors. Model routing includes automatic failover on error.

Model fallback middleware tries alternative models when the primary model fails, with provider redundancy across OpenAI, Anthropic and others.

Setting GOOGLE_CLOUD_AGENT_ENGINE_ENABLE_TELEMETRY enables agent traces, logs and metrics. Traces open in Cloud Trace and the console.

ADK emits OpenTelemetry spans for agent invocation, workflow execution, tool execution and model calls, exportable to Cloud Trace or an OTLP endpoint.

LangChain and LangGraph apps have a built-in integration that sends OpenTelemetry traces to LangSmith when LANGSMITH_OTEL_ENABLED is set.

Agent Runtime is a fully-managed runtime that abstracts away the underlying infrastructure, so you focus on agent logic instead of operations.

Google maintains and patches buildpack base images weekly; the page does not mention application dependencies. Instances are stateless, so state goes in external storage.

Google maintains and patches buildpack base images weekly; the page does not mention application dependencies. Instances are stateless, so state goes in external storage.

Data residency at rest is supported across Agent Runtime, evaluation, sessions and Memory Bank, alongside VPC Service Controls and CMEK.

DatabaseSessionService stores sessions in a relational database such as PostgreSQL, MySQL, MariaDB or SQLite that you manage yourself.

PostgresSaver stores checkpoints in a Postgres database reached through a connection string that you supply.

Agent Runtime bills vCPU-hours and GiB-hours of allocated compute and memory, rounded to the nearest second. Idle time between prompts is not billed.

Request-based billing charges for request processing, start and shutdown. Instance-based billing charges for the entire instance lifecycle.

Request-based billing charges for request processing, start and shutdown. Instance-based billing charges for the entire instance lifecycle.