Skip to content

AI Workforce & Assistance

Research agent: findings you can check

Some questions take an afternoon of tabs: permit records, manufacturer manuals, listings, your own files. A research agent works through the sources you approve, records every claim with its source and date, keeps disagreements side by side and says plainly what it couldn't find. It only reads and reports. A person judges the brief. We set it up with the sources you approve and the questions you ask most.
Talk through your researchSee the brief and its evidence ↓
Question · from W1Harbor Point Condos: what's installed, how old, and what needs servicing?County recordPermit portalPermit PDFAldmark manualListing · 2024Newsletter · 2021claimsourcestatus48 units, built 1987County recordsupportedBoilers replaced 2018Permit portalsupportedTwo Aldmark CB-800 boilersPermit PDFsupportedAnnual service for warrantyAldmark manual §7supportedHot water: in-unit or central?Listing vs newsletterconflictCurrent service contractornot foundbrief → W1 to review

Research is easy to start and hard to trust

Before W1 prices a maintenance plan for a 48-unit building, someone needs to know what's installed, how old it is and what the manufacturer requires. The answers are spread across a county record, a permit portal, a manual and an old newsletter, and the notes end up in someone's head or in a document nobody can check.

A research agent does the reading, but what it produces is evidence, not a verdict: each claim tied to the passage it came from, each gap admitted. That makes it quick to check and safe to build on.

Today, and with a research agent

Today

  1. 01Someone spends an afternoon across portals, PDFs and old emails.
  2. 02Notes go into a document, often without where they came from.
  3. 03Conflicting sources are settled by whichever was read last.
  4. 04What couldn't be found isn't written down at all.

The research can't be checked, so it gets redone, or trusted when it shouldn't be.

With a research agent

  1. 01The agent reads the approved sources for one question.
  2. 02Every claim goes into a table with its source and date.
  3. 03Disagreements are kept side by side, and gaps are listed.
  4. 04W1 reads a short brief, follows any citation, and decides.

Faster to produce, and quicker to check.

The brief, the evidence and the run

One question, three views of the same work.

Harbor Point Condos · 18 Harbor Way

for W1 · prepared Oct 22

  • The building has 48 units and was built in 1987 [1].
  • Its heating runs on two Aldmark CB-800 boilers, installed in 2018 under permit B-18-0412 [2][3].
  • The manufacturer requires annual service to keep the boiler warranty [4].

To settle on the site visit

  • Hot water: the 2024 listing says in-unit heaters [5]; a 2021 board newsletter says central [6].
  • Who services the boilers now: not found in any approved source.

Sources

  1. [1] County property record · checked Oct 22
  2. [2] City permit portal · B-18-0412
  3. [3] Permit attachment · equipment schedule
  4. [4] Aldmark CB-800 manual · §7
  5. [5] Sales listing · Mar 2024
  6. [6] Board newsletter · Jun 2021

The rules it works within

  1. Approved sources only

    A list, not the internet

    Specific sites, portals, files and records you approve. Anything else isn't read.

  2. Reads, never acts

    No forms, no messages

    It doesn't sign in, submit anything or contact anyone. Acting on the brief is for a person.

  3. Cite or drop

    Every claim has a passage

    A claim without a source doesn't go in the brief.

  4. Gaps are reported

    “Not found” is an answer

    What it couldn't find is listed, so nobody assumes it was checked.

  5. A budget for every run

    Time and steps

    Each run has limits, and stops with a partial brief that says what's missing.

The parts it's built from

Each part is built and tested on its own, so you can start with one and add the rest.

  1. Plan

    Tool-Using Agents

    Breaks the question into what has to be found.

  2. Search

    Semantic Search

    Approved sources, by meaning, inside access.

  3. Read and cite

    Retrieval-Augmented Generation (RAG) · LLM Application Development

    • passages
    • their source and date
    • the question

    no source → “not found”, never a guess

  4. Review

    Human-in-the-Loop AI · AI Evaluation & Guardrails

    A person judges; fixed questions test every change.

Underneath every stepAPI IntegrationConnects to the portals, files and records it may read.

What changes from one business to the next

SourcesWhich public sites, portals, files and internal records it may read
QuestionsStandard question sets for research you repeat, like a new commercial bid
OutputThe brief's format, and where it lands: the CRM, a shared folder or an email
FreshnessHow old a source may be before it's flagged

A common starting point is a hosted language model with tool calling, a search service and page reader limited to approved sites, PostgreSQL with pgvector for internal documents, and Langfuse for tracing. These are starting candidates, not requirements.

Is a research agent the right system?

Good fit

  • The same kind of research repeats: bids, new accounts, supplier checks.
  • The sources are known and can be listed.
  • A person will read the brief before acting on it.

Avoid when

  • The answer is in one internal document. An internal knowledge assistant answers faster.
  • You need a decision rather than evidence.
  • The sources sit behind sign-ins you can't give it read access to.

How we'd approach it

Most of the early work is choosing the sources and questions, the shape of the brief and how fresh findings must be.

  1. 01

    Understand

    As much as the problem needs

    We map how the work runs today: who does what, which systems hold the records, and where it breaks.

  2. 02

    Design

    Decided before it's built

    Ownership, rules, integrations and exceptions, agreed with you and written down with acceptance criteria.

  3. 03

    Build

    Connected to what you use

    We build the scoped system, connect it to the tools you already run, and test it on the cases that matter.

Related systems

The capabilities underneath

Questions about research agents

Does it search the whole internet?

Only the sources you approve: specific sites, portals, files and internal records. Anything outside that list isn't read.

What if sources disagree?

Both are kept, side by side, with their dates. The brief says they disagree and what would settle it.

Can it act on what it finds?

No. It reads and reports. Contacting anyone, submitting forms or updating records is for the person who reads the brief.

How long does a run take?

Minutes rather than seconds, because it reads many sources. Every run has a time and step budget and stops with a partial brief if it reaches it.

Get research you can check

Tell us which questions take your team an afternoon today. We'll show you what a brief with sources would look like.

Start a project