Agent teams

Donkit lets one agent draw on the help of others in several ways, from lightweight to fully published.

Sub-agents (helper agents)

An agent can spawn temporary sub-agents to tackle pieces of a larger task in parallel, then fold their results back into its own answer. This is a capability you enable in the agent's Tools (shown as Sub-agents). Sub-agents can themselves spawn helpers, up to a limited nesting depth, and each works against a scoped slice of memory so they don't interfere with one another. Use this when a single request naturally fans out into several independent chunks of work.

Connected agents (delegation)

You can connect an agent to your other agents so it can delegate to them. Mid-conversation, it calls a connected agent when that agent was built for exactly the job at hand, knows something it doesn't, or when you ask for it, and weaves the reply into its own response — one entry point for the user, a team of specialists behind it. You connect an agent from its Connected agents panel and optionally add a "When to call it" hint (for example, "Ask this agent anything about budgets and spending").

Beta

Connected-agents / multi-agent delegation is currently labeled Beta in the product.

In your own chat

You get the same thing without configuring anything: your main assistant already knows the agents you own and the ones shared with you. Mention agents with @ to have it delegate a message to them, or pin an agent to talk to it directly in the middle of the conversation.

Agent Teams (a published multi-agent product)

A Team bundles several of your agents (at least two) into one product, with one agent designated the Lead as the single chat entry point; members can call one another behind the scenes. Teams have their own shared team memory.

The Agent Teams heading in the left rail opens the teams page. The arrow collapses the list, which shows three teams. A row opens that team's chat. New team opens the form. There you name the team. Choosing the lead opens the list of agents who work behind that lead. You can publish a team to a public URL with the same end-user access controls as a published agent (private/invite, password, open, or organization-wide). Teams share the workspace balance and can be paused if a plan limit or credits run out.

Besides shared team memory, a team can have a shared team database — one set of tables the lead's Builder designs that every member agent reads and writes, with its own per-table access policy and per-user isolation for end users.

Turning it on

Sub-agents are a per-agent tool toggle. Connecting agents, building teams, and spawning helper agents are plan-gated — included in every paid plan — while existing teams stay manageable if you change plans.