andibase

Agents

What agents are in andibase, the main concepts behind them, and where they run

Open Markdown

Agents are reusable AI workers inside a workspace.

In andibase, an agent is not just a single chat response. It is a saved workspace resource with instructions, capabilities, and access to the shared context around your workflow.

You define an agent once, then run it in conversations where it can help users, use tools, and work against the same workspace data model as the rest of your team.

What makes up an agent

The core pieces of an agent in andibase are:

  • prompt
  • context
  • skills
  • tools

These pieces work together to shape how the agent behaves and what it can do.

Prompt

The prompt is the agent's standing instructions.

It defines what the agent is for, how it should respond, what kinds of tasks it should handle, and any constraints you want it to follow. In practice, this is the long-form instruction set you save on the agent definition.

Examples:

  • a support agent that answers concisely and escalates billing issues
  • an operations agent that updates records after each completed task
  • a research agent that summarizes findings before taking action

The prompt is the baseline behavior. It stays consistent across chats unless you edit the agent.

Context

Context is the information available to the agent at runtime.

This includes the current conversation, the workspace the agent is running in, and any relevant data the runtime can resolve for that chat. Context is what lets the same agent behave differently in different situations without changing its base prompt.

In andibase, workspace context matters because it controls what data and resources the agent can actually access.

Skills

Skills are reusable instruction bundles that help an agent perform a specific kind of work well.

A skill can encode a workflow, a domain-specific pattern, or a preferred way to use tools. Instead of repeating those instructions in every prompt, you can treat them as a focused capability layer the agent can rely on.

Use skills when you want:

  • the same workflow reused across multiple agents
  • a standard operating procedure for a task
  • a clearer separation between general behavior and specialized execution patterns

Tools

Tools are the actions an agent can take beyond generating text.

They let the agent read from or write to workspace resources in structured ways. Tool access is controlled by the authenticated caller and the resolved workspace context for the chat.

Current runtime tools are focused on workspace data access, including:

  • listing data definitions
  • querying records
  • creating records
  • updating records
  • deleting records

See Agent Tools for the current runtime surface.

Where agents run

Agents in andibase currently run in workspace-scoped chat experiences and API-driven runtimes.

Workspace chats

You can run an agent from the workspace UI, where each agent has its own chat threads inside a workspace. This is the main product surface for interacting with agents as a user or team.

Programmatic API flows

You can also run agents through the HTTP API after authenticating as an agent or other authorized caller. This is useful when you want an external system or another runtime to start or continue agent work.

See Get started (for AI Agents) for the current authentication and API flow.

Workspace-aware tool runtime

When an agent runs with a resolved workspace context, it can use the runtime tools available for that caller. This is what turns an agent from a pure chat interface into a workspace operator.

Putting it together

An agent in andibase is best understood as:

  • a saved prompt
  • running with live context
  • optionally guided by skills
  • able to use tools when permissions allow

That combination is what makes agents useful for repeatable, real workflows instead of one-off chats.

See also:

On this page