> ## Documentation Index
> Fetch the complete documentation index at: https://docs.within.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Build an agent with the Within ADK

> Go from a roadmap item to an agent running in production: specify, build and validate it with the Within ADK, in the coding agent you already use.

The Within Agent Development Kit (ADK) takes an item from your roadmap, and the objective behind it, to an agent running in production. The agent is built from your own observed workflows rather than a generic template, and the work happens in the coding agent your team already uses.

> [!NOTE]
>
> **Before you start:**
>
> * The Within MCP connected to your coding agent. See [Using the Within MCP](/user-docs/deliver/within-mcp).
> * A use case chosen in Define, with its objective and constraints. See [Define your solution set](/user-docs/define/define-your-solution-set).
> * Work Brain coverage of the process the agent will work in.

## How the ADK works

There's no separate package to install. Once Within is connected, your coding agent fetches the current method from Within at the start of each build and follows it, so every build uses the latest version.

Keep the agent's code, configuration, evaluations and evidence baseline in your own repository, under version control.

## Specify, build, validate

**1. Specify**

* **Objective.** Restate the business objective and the technology stack the agent has to fit.
* **Coverage.** Confirm your Work Brain covers the area. Where it's thin, run Companion there before you build.
* **Current state.** Read the observed work: pain points, bottlenecks, time and motion, the systems in use, and the bright spots.
* **Use case.** Screen candidates on evidence, expected impact, feasibility and the controls they need, and pick one.
* **Design.** Write the agent specification, effectively a product spec for the agent: scope, governance, credentials, flow, and where a person approves.

A person approves the design before anything is built.

**2. Build**

Generate the agent's code, skills and configuration against the specification.

**3. Validate**

Generate and run evaluations before release, against approved test data, with evidence for every result: ordinary cases, edge cases, malformed or missing inputs, and requests the agent should refuse. A person approves the release. See [Validate an agent before release](/user-docs/deliver/validate-an-agent).

Then deploy it into your environment, for example as a managed agent.

Along the way you get a current-state read and a recommendation you can share with stakeholders, the agent specification, and the code, skills, configuration and evaluation results.

## What you can build

Choose the smallest shape that produces the outcome and passes its controls. A scheduled trigger doesn't on its own mean you need an agent.

| Shape | What it is | When it fits |
| - | - | - |
| **Skill** | An instruction package for one bounded task | A single task small enough to review directly |
| **Integration Flow** | Rule-based automation on a fixed, predictable, auditable path. It doesn't use AI reasoning, so it isn't an agent. | The process is stable and every step must be explainable |
| **Managed agent** | An AI agent that plans its own path under a person's approval. Chained across several tasks and tools, it becomes a compound agent. | The inputs vary too much for a fixed path |
| **Context bundle** | Packaged context from your Work Brain that grounds another AI tool or a code-generation build | Another tool needs Within's evidence to do its job |

## Which tool can I build in?

| Who's building | Where | What they build |
| - | - | - |
| Developers and AI platform teams | Coding agents such as Claude Code, Codex, Cursor, Gemini CLI or GitHub Copilot, connected to the Within MCP | Compound agents and Integration Flows, with the full ADK sequence |
| Business users | Advisor, or a chat app such as Claude or ChatGPT connected to the Within MCP | Skills and simple agents, one bounded task at a time |

Install guides for each client, including Microsoft 365 Copilot, are in the [Within MCP developer docs](https://developers.within.ai/quickstart).

## Controls that stay on

* Every material claim about the process cites observed data.
* A person approves the design and the release.
* Evidence from your Work Brain doesn't replace security review, system documentation or runtime testing.
* Connecting the Within MCP doesn't authorize actions in other systems. Set write access and approvals in your AI client or in the destination system.

## Getting help

Within's team can work with your engineers on implementation, production readiness and evaluation design. Ask your Within team.

## Where to go next

> **[Validate an agent before release](/user-docs/deliver/validate-an-agent)**
>
> Test it, pilot it, and decide whether it's ready.

> **[Using the Within MCP](/user-docs/deliver/within-mcp)**
>
> Connect Within to your coding agent or chat app.

> **[The technical track (developer docs)](https://developers.within.ai/transformation/technical-track)**
>
> The full build sequence for developers.

> **[The non-technical track (developer docs)](https://developers.within.ai/transformation/non-technical-track)**
>
> Skills and simple agents for business users.

> [!NOTE]
>
> **Need help?** Use the in-app chat — click the chat bubble in the bottom-right corner of Within (staffed 24×5) — or email [support@within.ai](mailto:support@within.ai).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.