Building QuikSignal: What Running Eleven AI Research Agents Taught Us

Most of our work is building AI systems for clients. QuikSignal is one we built and run for ourselves: an AI market and competitive intelligence platform in which eleven AI agents research the companies and markets a team chooses, cross-check what they find, and write a short brief where every line opens the document behind it.

Running it every day taught us more than any client demo could, because there is nobody to hand it over to. These are the lessons we now bring to client projects.

Give each agent one job

Our first instinct was a single capable agent with a long prompt. It was hard to test and harder to debug. QuikSignal now runs eleven agents, each with one responsibility: Scout researches each company, Pricer follows pricing and packaging over time, Recruiter reads hiring, Regulator reads filings and upcoming rules, Historian reads ninety days of history first, Sifter decides what matters, Mapper keeps decision-makers current, Verifier cross-checks findings, Forecaster projects series and scores its own record, Analyst writes the brief, and Router delivers it.

Narrow roles make each step testable on its own, and when an output is wrong you can see which step produced it. The same principle applies to most multi-step GenAI systems we build for clients.

Make every claim carry its evidence

A summary that cannot be checked is hard to trust. Every QuikSignal finding opens the dated document it came from, and shows how many independent sources carry the same fact. We are careful with that count: agreement between sources is reported as agreement, not as proof.

For client systems this usually means retrieval with citations: the answer and the passage it was built from, side by side. It adds engineering work, and it is the single feature users mention most when they decide whether to rely on a system.

Log what you hold back

Filtering is as important as finding. QuikSignal holds back items that change nothing, merges near-duplicate coverage of one story, and writes every held-back item to a log with the reason. Users can open that log at any time.

The log turned out to be a trust feature. When someone asks "why didn't it mention this?", the answer is one click away instead of a support ticket.

Precision beats coverage

An early version matched government contract awards to companies by similar names. It produced plenty of findings, and a review showed that many of them belonged to a different company with a similar name. We switched to exact matching on the recipient's legal name and withdrew the findings that no longer qualified.

The product showed fewer findings afterwards, and it was better for it. One wrong finding costs more trust than ten missing ones. When we scope client systems, we now agree the acceptable error rate before agreeing the coverage.

Plan for the model provider failing

Model providers have outages, rate limits and price changes. QuikSignal sends its model calls through one internal router with several providers behind it, so a failing provider is skipped and the work continues. Switching a task to a different model is a configuration change, not a rewrite.

We recommend the same layer to every client: it costs little to build at the start and a great deal to add later.

Read history before writing anything new

A finding only matters relative to what came before. Before writing, QuikSignal reads ninety days of history for each company, so a price change is compared with earlier captures and a hiring spike with the usual pattern. Without that context, an AI system will happily report old news as new.

Test the words, not just the code

Language is part of the product. We run automated checks over every customer-facing string for wording we have decided never to use. Two lessons came out of it. First, a test can lock a mistake in place: one of ours was asserting that a sentence we later wanted gone must exist. Second, a check that reads source code can miss text split across two concatenated strings, so the check now joins them before matching.

Keep cadences fixed

Some of QuikSignal's research runs on a fixed weekly schedule so that one week can be compared with the next. It is tempting to trigger an extra run when you want fresh numbers. We don't, because an irregular cadence makes the comparison meaningless. Reading the schedule from the code before concluding a job did not run saved us more than one false alarm.

What this means for client projects

Each lesson above now shows up in how we scope GenAI work: narrow agent roles, citations on every answer, a log of what the system chose not to say, error budgets agreed up front, a provider-neutral model layer, and tests that cover wording as well as code. If you are planning an AI agent system, our Data and AI engineering team can help, and you can see QuikSignal itself at signal.quiksync.com.