Can the agents make up information?

No system eliminates this completely. The platform reduces it in three ways: every claim must cite real excerpts from your corpus, an automatic validation checks it and, without enough evidence, the answer is “not enough data”. That is why each citation links to its source and shows the excerpt: you can verify it.

Which plans include the agents?

All three plans include the six agents. Those that use an external AI provider require the organization’s administrator to allow it. The demo shows how they behave with sample data and prerecorded answers.

What does “not enough data” mean?

That there is no evidence among the collected items to answer. The agent says so instead of filling in the answer with assumptions, and states the reason: nothing was found about it, what was found does not answer the question, or the answer could not be supported with valid citations.

AI agents — Atalaya

What each agent does

Tag generator: Classifies each item with tags that match the channel’s topics.
Key message generator: Writes up the essential points of a channel over a period, from most to least important, each with its sources.
Summarizer: Summarizes a piece of content or a story covered by several sources, and states how many pieces of content it is based on.
Content recommender: Ranks what you have not yet rated according to what you marked as “interesting”, and shows why it recommends each item.
Source recommender: Suggests sources you could add: those in the catalog you do not have yet and the sites that already appear in your items. It scores them by the evidence that supports them, their relevance, your ratings and their presence in the catalog. It does not look anything up online.
Questions to the corpus: Answers natural-language questions about what has been collected, with at least one linked citation in every answer. You can scope by channel and dates, and look up the history.

How citations are verified

The agent receives excerpts from your items, each with a reference that is valid only for that call; it never sees internal identifiers.
It must return each claim together with the references of the excerpts that support it.
An automatic validation checks that each reference corresponds to an excerpt that was provided. An unknown reference invalidates the entire claim.
If the validation fails, it is retried once. If it fails again, the summary answers “not enough data”; key messages and answers discard only the invalid claim and, if none is left, they also answer “not enough data”.
Each citation is stored with the excerpt that supports it, so you can check it.

Every claim links to its source. If there is no evidence, we say so.

Safeguards

Source content is data, never instructions: Captured text is delimited and sanitized. The agents do not take actions with side effects based on what a page says.
Versioned instructions: Each run stores the model, the version of the instructions and the sources used.
Evaluation before every change: Every change of model or instructions is measured against reference sets for relevance, tagging, summaries with citations and questions.
Customer data: An organization’s content is not sent to an external AI provider unless its administrator allows it. The permission is per organization and is turned off by default.
Usage and budget in plain view: Every call to a model is recorded in a run log. An organization can have a cap on AI usage; once it is reached, the agents stop calling the model.
Interchangeable models: An abstraction layer makes it possible to use a local model or a commercial API, depending on the configuration.

Human review is part of the process

Each person can rate an item as “interesting” or “not interesting”. That rating is personal and feeds the recommenders. In addition, the organization’s ratings as a whole fine-tune the relevance of similar items scored afterward, within a bounded margin and with the explanation in plain view. And newsletters go through a person, who reviews and edits them before they are sent.

See how they behave in the demo

Frequently asked questions

Can the agents make up information?: No system eliminates this completely. The platform reduces it in three ways: every claim must cite real excerpts from your corpus, an automatic validation checks it and, without enough evidence, the answer is “not enough data”. That is why each citation links to its source and shows the excerpt: you can verify it.
Which plans include the agents?: All three plans include the six agents. Those that use an external AI provider require the organization’s administrator to allow it. The demo shows how they behave with sample data and prerecorded answers.
What does “not enough data” mean?: That there is no evidence among the collected items to answer. The agent says so instead of filling in the answer with assumptions, and states the reason: nothing was found about it, what was found does not answer the question, or the answer could not be supported with valid citations.

Start monitoring your sector today

Create the account, choose your sector and in a few minutes you will have sources bringing in information, a channel and your first watchlist.

Create a free account

Questions? Write to us.