Agents

Agents are specialist sub-agents that Pandora can route tasks to. When a user asks a question that matches an agent’s expertise, the main agent automatically delegates — no routing configuration needed.

Unlike tools, agents don’t contain executable code. They are pure data: an identity (name, description) and instructions (system prompt). The system handles execution, model selection, and tool access.

Agent Definition

Each agent entry point exports a named agent object:


import type { Agent } from '@pandorakit/sdk/agents'
 
export const agent: Agent = {
  id: 'research',
  name: 'Deep Research',
  description: 'Thorough, multi-step research on complex topics',
  instructions: `You are a deep research specialist. Your job is to thoroughly
investigate complex questions by conducting multiple rounds of web searches,
cross-referencing sources, and synthesizing your findings.
 
When answering:
- Cite all sources with URLs
- Distinguish between well-established facts and claims with limited sourcing
- Note any conflicting information you found`,
}

Manifest Fields

The provides.agents section of your manifest accepts an array of entries:


{
  "provides": {
    "agents": [
      {
        "entry": "./src/research.ts",
        "useTools": ["@pandorakit/tavily-search:tavily_search"],
        "modelTools": ["search"]
      }
    ]
  }
}

Field	Description
`entry`	Path to the file that exports the agent definition
`useTools`	Tool IDs from the global registry. Use fully-qualified IDs: `pluginId:toolId`
`modelTools`	Model-native capability keys (e.g. `"search"` for provider’s built-in web search)

Enable It

Users control agents from the Plugins page — toggling the plugin on/off and optionally configuring per-agent model overrides.

Routing

Pandora reads each agent’s description and decides whether to handle a request directly or hand it off to a specialist. The routing decision is made by the LLM, not by rules — adding a new specialist is just installing a plugin.

With zero agents registered, Pandora handles all requests directly — behavior is unchanged from a standard setup.

The quality of your agent’s description directly affects how well routing works:


// Good — specific about what it does
description: 'Thorough, multi-step research on complex topics using web search'
 
// Bad — too vague
description: 'A helpful assistant'

Include the kinds of tasks the agent handles and the tools it has access to (conceptually, not by ID).

Model Overrides

Users can configure a per-agent model override on the Plugins page. This lets them use a different (potentially cheaper or faster) model for specific specialists.

Reference


import type { Agent } from '@pandorakit/sdk/agents'

`Agent`

Field	Type	Required	Description
`id`	`string`	Yes	Unique agent identifier
`name`	`string`	Yes	Human-readable display name
`description`	`string`	Yes	What the agent does (used for routing decisions)
`instructions`	`string`	Yes	System instructions for the agent
`useTools`	`string[]`	No	Tool IDs from the global registry (set from manifest at load time)
`modelTools`	`string[]`	No	Model-native tool keys like `'search'` (set from manifest at load time)

useTools and modelTools are typically declared in the manifest rather than in the agent code.