> ## Documentation Index
> Fetch the complete documentation index at: https://docs.akua.dev/llms.txt
> Use this file to discover all available pages before exploring further.
# Configure an agent
> Define an agent with skills, model policy, runtime policy, grants, and optional ambient triggers
Create an agent when you want a reusable automation identity for a workspace. A production triage agent, a repository-change agent, and a seller support agent can all use the same resource model with different instructions and policies.
Agent resources are available through the API. Dashboard configuration screens will expose the same fields as product UI.
## What you configure
| Setting | Purpose |
| ------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Name and description** | Helps users choose the right agent. |
| **Instructions** | Defines the agent's role and operating style. |
| **Skills** | Adds reusable behavior such as repository change triage or deployment investigation. |
| **Capabilities** | Enables broad modes such as chat, coding, and ambient operation. |
| **Communication style** | Controls whether the agent explains work in beginner, intermediate, advanced, expert, or adaptive language. See [Communication style](/agents/communication). |
| **Model policy** | Controls model selection, budget, and whether provider billing is Akua-managed or workspace BYOK. |
| **Runtime policy** | Defines when the agent can use a retained runtime, filesystem, shell, or browser. |
| **Grants** | Limits which platform resources the agent can read or mutate. |
| **Permission policy** | Decides which actions require human approval. |
| **Ambient triggers** | Starts work automatically when configured signals match. |
## Grants
Grants are explicit permissions. Avoid giving an agent broad mutation access when it only needs context.
For agents that prepare repository change requests, the recommended baseline is:
| Resource | Actions | Why |
| --------------------------- | ---------------------------- | -------------------------------------------------------------------------------------------------------------- |
| `repository_change_request` | `read`, `create`, `continue` | Read existing change requests, create a reviewable draft, or continue a prior change request when appropriate. |
| `install` | `read` | Inspect installation state and source configuration. |
| `cluster` | `read` | Gather operational context. |
| `product` | `read` | Understand the packaged software and versions. |
Additional actions such as update, close, or accept should be separate grants and can require approval.
## Model policy
Agents can use Akua-managed billing or workspace-owned provider credentials. Workspace credentials are referenced through secrets; secret values are not exposed to the agent runtime.
Use budget limits for agents that can run automatically or launch coding work.
## Communication style
Agents should match the user's experience level. A beginner may need "your app keeps restarting" with a suggested next step. An expert may prefer exact resource names, recent events, and the underlying Kubernetes status.
Use an adaptive style when the agent should vary detail within an explicit user, session, or agent default. The agent should not silently change a user's proficiency level based on writing style.
Communication style should not reduce analysis quality. It changes how findings are explained, not what evidence the agent gathers or records.
## Runtime policy
Not every task needs a retained runtime. Prefer the cheapest safe runtime:
| Runtime need | Use |
| --------------------------------- | --------------------------------------- |
| Read platform state and summarize | No retained runtime |
| Call Akua API through tools | No retained runtime |
| Edit a repository or run tests | Retained filesystem |
| Run commands | Shell, usually approval-gated |
| Inspect a web app visually | Browser, usually disabled unless needed |
## Related topics
Configure background triggers and policy.
Choose safe grants and approval gates.