Skip to content
MAIA LogoDocs

Capabilities/Agents

Sub-agents

MAIA is a coordinated system of agents. The main agent runs the conversation, and when a question needs research on every row, it hands the tasks to a sub-agent. Hundreds run at once, in the background, and none of them holds up the main agent.

One conversation, many researchers#

You talk to one agent. Behind it is a system. When your objective needs a fact the data does not hold, for every place on a list, the main agent does not research the list itself. It writes the question once, hands the tasks to a sub-agent, and goes on with the conversation.

Sub-agents are the main agent's research tool. Each takes one row and one question, goes as deep as that row needs, and returns an answer with the evidence for it. Hundreds run at once, and their work lands in the table as it finishes.

Built for the job#

MAIA's sub-agents are not a general chatbot pointed at a row. We built them for this work and fine-tuned them on it, so each one knows how to read a parcel record, a company filing, or an overhead image, and when to stop. They are token-efficient by design: a sub-agent spends what the row needs and no more. That is what makes research on ten thousand rows practical, in time and in credits, and you will not find it elsewhere.

What gets delegated#

Anything that has to be found out about one place at a time.

Is this roof ready for solar?What the sub-agent doesReads the overhead image of that roof
Does the owner occupy this building?What the sub-agent doesCompares the owner with the occupant, in the records and on the web
Who runs this business, and how do I reach them?What the sub-agent doesResolves the entity and finds the person
Has this company published an ESG commitment?What the sub-agent doesSearches the web and reads what it finds

The main agent phrases the question, tests it on a few rows, and corrects it before it dispatches the rest. The kinds of question MAIA can research are on Enrichments.

What comes back#

Every sub-agent returns the same three things, in one cell.

  • A value, typed as yes or no, a number, a category, a link, or text.
  • The reasoning, in one or two sentences that lead with the finding.
  • The sources the answer rests on.

When a sub-agent finds nothing it can support, it says what it checked and returns no value. The cell reads Unavailable. It never fills the gap with a guess.

Four questions, answered row by row. Unavailable is where nothing supported an answer.

Nothing waits on them#

Sub-agents run on MAIA's servers, in parallel, and the main agent does not block on them. While rows fill in you can keep talking to it, open the map, or close the project and come back. The rows are still there, and so is the answer.

The one thing that waits is the main agent's answer. It reads the finished cells and checks the result before it quotes a value. See Main agent.

What MAIA checks#

An answer from web research is screened before it is saved. One with no research behind it is rejected: the cell reads Unavailable and the row is not charged. An honest "could not determine" passes. The screen is for research that was never done, not for answers you would disagree with.

Limits#

  • No automatic retry. A run skips Unavailable rows. Run them again with Re-run unavailable rows in the column menu, which spends credits.
  • Each question has a research budget. A sub-agent stops when it reaches it, so a hard row comes back Unavailable rather than running on.

Related#