Support dynamically defined agents created at runtime

[schplitt]

RFC

Summary

Allow Flue applications to define and run agents from runtime config instead of only from statically authored 'use agent' modules.

Background & Motivation

Today, Flue v2 appears to be built around statically defined agent functions discovered from source files. That works well for code-first applications, but it makes it hard to build products where users can create and run agents directly from the frontend.

I would like to build a UI where a user can create an agent by choosing a model, provider, instructions, skills, tools, sandbox settings, and other options from a simple config form. That config would be stored in a database and then used to run the agent without a redeploy.

Right now it seems possible to make behavior dynamic inside a fixed agent function, but not to define truly dynamic agents as first class runtime agents. This makes product use cases harder, especially for multi-tenant apps, internal agent builders, and no-code or low-code workflows.

Goals

• Allow applications to define agents from runtime config
• Allow those agents to be created and run without redeploying
• Support dynamic configuration of model, provider, instructions, tools, skills, MCP connections, and sandbox settings
• Make this work for deployed applications, not only local scripts
• Preserve Flue's existing durability and runtime guarantees as much as possible
• Enable frontend driven agent creation for product use cases

Example

A simplified example of the kind of API that would help:

const agent = defineDynamicAgent({
  name: 'customer-support-helper',
  model: 'anthropic/claude-sonnet-4-6',
  provider: 'anthropic',
  instructions: 'Help the customer solve support issues.',
  tools: ['lookup_order', 'create_ticket'],
  skills: ['refund-policy', 'tone-guide'],
  sandbox: config.sandbox === "local" ? local() : daytona(),
});

const handle = init(agent, { id: 'conversation-123' });
const receipt = await handle.dispatch('Help this customer with their order issue.');
const reply = await handle.read(receipt);

Or even a lower level API such as:

const agent = createAgentFromConfig(configFromDatabase);

Example use case from a frontend product:

• User opens an app
• User creates a new agent in a UI
• User selects model, tools, skills, and prompt settings
• App stores that config in a database
• App runs that agent immediately without a rebuild or redeploy
• Future edits to that config affect future runs

This would make Flue much easier to use as the backend for user-defined agent products.