The short answer
An agent development kit (ADK) is a packaged set of building blocks for creating, testing and running AI agents: a model interface, tools that act on systems, orchestration, guardrails, human checkpoints, evaluation, tracing and a runtime. Examples include Google ADK, the OpenAI Agents SDK and the Claude Agent SDK. Most are built for engineers.
What's in the box
Strip away the branding and every kit answers the same ten questions. Any piece a kit leaves out becomes code your team writes and maintains.
Model interface
What it doesTalks to the model
Ask thisCan I swap the model for one step?
Tools and actions
What it doesReads and changes systems
Ask thisWhat can it touch, and what's fenced off?
Orchestration
What it doesDecides what runs next
Ask thisWhere is the path fixed, where does the model choose?
State and memory
What it doesCarries values between steps
Ask thisWhere are values stored, and for how long?
Rules and guardrails
What it doesValidates and blocks actions
Ask thisWhich decisions are exact rules?
Human checkpoints
What it doesPauses for a person
Ask thisDoes it wait, or time out and continue?
Evaluation
What it doesRuns test cases
Ask thisCan I rerun a fixed test set after every change?
Tracing
What it doesRecords each step
Ask thisCould an auditor rebuild one run?
Runtime
What it doesRuns the agent
Ask thisWhose machine, and where does data go?
Identity and secrets
What it doesHolds credentials
Ask thisDoes the model ever see a password?
Workflows vs agents
A workflow runs models and tools along paths written in advance. An agent lets the model choose its own path. For back-office work the choice is rarely close: invoice posting, pay-app review and rent reconciliation have known steps and known exceptions. They're workflows with model calls inside bounded steps. Open-ended research is where model-directed loops fit.
SDK, framework, platform: the labels blur
Vendors use these words loosely. Judge by what ships.
| Type | What you get | Who runs it | Examples |
|---|---|---|---|
| SDK or library | Code packages for an agent loop in your own program | Your engineers | OpenAI Agents SDK, Claude Agent SDK |
| Framework | Structure and tooling for multi-step and multi-agent apps | Your engineers | Google ADK, Microsoft Agent Framework, LangGraph |
| Managed platform | Hosted runtime, sessions, memory, evaluation | A cloud vendor, billed by use | Gemini Enterprise Agent Platform, Anthropic Managed Agents |
They stack. A managed platform often runs agents written with a framework, which sits on an SDK.
How to evaluate a kit
Ask these in order. The early ones rule out more options.
Answer the first three questions about your team.
Your resultSDKEngineering team, systems with good APIs: an SDK or framework.
Your resultFrameworkEngineering team, systems with good APIs: an SDK or framework.
Your resultManaged platformCompany standardized on one cloud: that vendor's managed platform.
Your resultReady-to-run studioBack office on legacy ERPs, portals and PDFs, no developers to spare: skip the kit and use a finished tool that works through the screen.
Answer all three to see a result.
What can it reach: APIs, a browser, or any desktop app?
Where are exact values decided: code, a prompt, or readable rules?
How does a person step in: a pop-up, or a real review queue?
What evidence exists per step?
Where does it run, and what leaves the network?
What happens on failure?
How are versions approved?
What can you export if you leave?
If a kit answers questions 1 and 2 badly for your team, the rest don't matter. A code-first framework in a company with no developers is shelfware.
Who each option suits
Engineering team, systems with good APIs
an SDK or framework.
Company standardized on one cloud
that vendor's managed platform.
Back office on legacy ERPs, portals and PDFs, no developers to spare
skip the kit and use a finished tool that works through the screen.
Many companies need two: engineers build customer-facing agents on a framework, operations builds process agents in a ready-made tool.
Working agents without a kit
LaunchAI Studio is for teams that want agents doing real work without building on a dev kit. There's nothing to assemble. The person who owns the process talks to the Studio and shows their screen, the Studio proves each step live, and the agent runs on your own machine across any app. Product →
What no tool supplies
Three things come from the business, whatever you choose: a written description of the process including the rare exceptions, the exact rules and thresholds, and a set of real cases with known right answers. Agent-Ready SOP →
- 1A written description of the process, rare exceptions included
- 2The exact rules and thresholds
- 3A set of real cases with known right answers
Questions
Do I need a kit to use AI agents?
No. Many business apps now ship agents inside their own screens. A kit or tool matters when work crosses apps no single vendor controls.
Can a non-developer use a code-first kit?
Partly. Some accept configuration files instead of code, but wiring, deployment and debugging still assume an engineer.
What should I be able to export?
Step definitions, rules and tables, version history, test cases and run evidence, in a readable format.
Keep learning
- What Is a Computer-Use Agent A computer-use agent is AI software that operates a computer through its screen. Read the guide
- Automate Without an API First figure out which "no API" you have. Read the guide
- Which Workflow First Start with a workflow that is frequent, slow by hand, low in variation, spread across a few systems, cheap to get wrong and rarely off the happy path. Read the guide
No prompts. No babysitting. No Integrations.