What is Agent Development Kit?
Agent Development Kit is a software framework that packages reusable interfaces and runtime services for defining, executing, observing, evaluating, and deploying AI agents.
How It Works
Definition layer and runtime layer
The definition layer describes an agent's name, instructions, model interface, tools, output contract, and optional child agents. The runtime layer owns turns, tool-call loops, event delivery, state persistence, cancellation, limits, interruption, resume, and finalization. Keep these layers separate so an agent definition can be evaluated with a fake model and the runtime can enforce policy independently of model instructions.
Agents versus deterministic workflows
Use an agent when the next action is genuinely open-ended and model judgment is useful. Use ordinary code or a workflow graph when steps, branches, or completion rules are known. A capable ADK should compose both rather than forcing every node to be an LLM agent. Multi-agent support, memory, and planning are optional capabilities, not evidence that an application needs them.
Selection and portability contract
Compare message and event schemas, synchronous and streaming APIs, tool semantics, state scopes, checkpoint compatibility, human interruption, observability, evaluation, deployment targets, language support, and version policy. Provider-neutral interfaces still leak capabilities such as reasoning blocks, strict tool schemas, multimodal parts, server-side tools, and finish reasons. Build a thin application-owned boundary and test migrations with recorded fixtures instead of assuming drop-in replacement.
Security and production responsibility
An ADK can expose guardrails, callbacks, approvals, and tracing, but it cannot decide an application's authorization or make external effects idempotent. Enforce tool allowlists, resource-level access, argument validation, budgets, deadlines, audit records, and approval in application code. Checkpoint logical progress, reconcile uncertain side effects, redact telemetry, and pin the framework, model adapter, prompts, tool schemas, and evaluation suite in each release.
Key Characteristics
- Defines agent, model, tool, message, event, and output interfaces
- Runs bounded turns with cancellation, streaming, errors, and terminal states
- May provide session state, memory, artifacts, checkpoints, and resume
- Composes model-driven agents with deterministic workflows and human input
- Exposes callbacks or middleware for tracing, evaluation, and policy hooks
- Leaves business authorization, effect safety, and release accountability to the application
Common Use Cases
- Building a bounded tool-using assistant with typed messages and outputs
- Combining model decisions with deterministic workflow and approval nodes
- Persisting and resuming long-running agent work through checkpoints
- Standardizing model adapters, tools, sessions, events, and telemetry across teams
- Testing an agent runtime with fixtures before controlled production rollout
Example
Loading code...Frequently Asked Questions
Is Agent Development Kit a standard or a specific product?
It is a framework category and also appears in specific product names, especially Google ADK and Eino ADK. There is no universal ADK interface. Each kit defines its own agent lifecycle, state, events, tools, checkpoints, and deployment model, so identify the implementation and version whenever behavior matters.
What should a good ADK make explicit?
It should expose model and tool contracts, state scopes, event and streaming semantics, limits, errors, terminal states, interruption, approval, checkpoint and resume, telemetry, and evaluation. Hidden retries, implicit context mutation, or unclear side effects make an agent difficult to reproduce and govern.
When should an ADK workflow be used instead of an agent?
Use deterministic workflow nodes when order, branching, policy, or completion can be expressed in code. Use a model-driven agent only where open-ended judgment adds measurable value. Most production applications combine both, keeping authorization and irreversible effects outside the model's discretion.
Does using an ADK make an agent production-ready?
No. The framework can provide runtime primitives, but the application still owns data governance, resource authorization, tool validation, idempotency, reconciliation, evaluation thresholds, monitoring, incident response, and rollback. Built-in guardrails are useful controls, not proof of safety.
How should teams compare Agent Development Kits?
Implement the same representative workflow and failure cases in each candidate. Compare lifecycle clarity, capability fit, portability, state and checkpoint behavior, streaming correctness, observability, security hooks, testability, deployment ownership, upgrade policy, latency, and accepted-outcome cost rather than feature-list size.