Category

AI research agents

An AI research agent is a program that pursues a research goal across several steps - choosing sources, fetching them, extracting what matters and assembling a result - rather than answering from memory in one shot. The useful ones are narrow and specialised: each agent has one job, one set of sources and a defined output, which is what makes their work checkable.

Agent is doing a lot of work in that sentence

The word has been stretched to cover everything from a prompt template to an autonomous system. A working definition that stays useful: an agent takes a goal rather than a question, decides what to do next based on what it has found so far, uses tools to act on the world, and stops when the goal is met or a budget is spent.

By that definition, a chat interface answering from its training data is not an agent. A program that reads six job boards, notices which companies posted new roles, and writes them into a record is, even though it contains no clever reasoning at all.

Why several narrow agents beat one broad one

A single general agent asked to "research this company" produces output whose quality nobody can predict, because nobody can say which sources it happened to read. Splitting the work into specialists changes that: one agent reads job boards, another reads pricing captures, another reads filings, another decides what is worth keeping.

The benefits are practical rather than theoretical. Each agent can be tested against its own sources. A failing source degrades one agent rather than the whole run. And the output can name which agent produced each finding, which is the difference between a report you can audit and a report you have to trust.

The failure modes to check for

Three recur across every multi-agent research system, and all three are quiet:

  • Fabricated citation. The model echoes back a URL that looks right and does not resolve, or resolves to something that does not say what was claimed. The fix is mechanical: fetch the source and bind the claim to it before the model is involved, rather than asking the model to remember it.
  • Duplicate counting. The same event arrives from five sources and is reported as five findings, which makes a quiet week look busy.
  • Silent source failure. A source goes down or changes shape, the agent that depends on it returns nothing, and the report simply has less in it. Without per-source reporting, nobody notices for weeks.

What to ask of any agent system

Which agent produced this finding, which source did it read, and what did the other agents fail to do on this run. A system that can answer all three is one you can debug. A system that presents a single confident narrative with no seams is one where you will find out about a broken source when a customer corrects you.

Where QuikSignal fits

QuikSignal is built this way: 11 named agents, each with its own sources and its own job - reading job boards, comparing pricing captures, reading filings and registers, scoring what is found, checking corroboration, and writing the brief.

Every run reports per source what answered, what was skipped because the company is not listed there, and what genuinely failed, so a silent source failure shows up as a row rather than as a thinner report.

See it on your own market →

What it does not do
  • The agents do not browse arbitrarily. Each reads a defined set of named sources, which is a deliberate constraint: it makes the output reproducible.
  • No agent makes a decision on your behalf, sends anything to anyone, or takes an action outside the product.
  • A company that says nothing in the open produces little. The system reports the gap rather than filling it.
Questions

What people ask

How many agents should a system have?
However many distinct jobs there are. The number is an implementation detail; what matters is whether each one has a defined source set and a checkable output. A system with twenty vaguely defined agents is harder to trust than one with four clear ones.
Do research agents need a large language model for every step?
No, and the better designs use one sparingly. Collection, deduplication and scoring are deterministic work that code does more cheaply and more reproducibly. The model earns its place at the end, on the small amount of material that survived.
What happens when two agents disagree?
In a good system, the disagreement is surfaced rather than resolved silently. Two sources contradicting each other is information; a system that quietly picks one has thrown that information away.

Read your own market the same way.

Eleven agents, the companies you choose, every night, with the document behind every line.