The Web Researcher

When the answer isn't in your memory, Previously doesn't make one up, and it doesn't dump a screen of blue links on you. It hands the question to the colleague who knows how to research — and passes back that colleague's answer, along with how sure they are.

webSearch is not a wrapper around a search box. The main agent hands it a research question in natural language — "find out whether this medication actually conflicts with coffee" — the way you'd assign work to a colleague who knows their way around sources. What comes back isn't ten links; it's an answer written after reading the material. You didn't ask it to "search" — you asked it to figure something out.

How it researches

Two steps, both in its own hands:

The search itself is provider-executed, producing the candidate material.
It reads pages itself. Its own webFetch opens the most promising pages for a real read — at most 3 pages, quota-bounded — because a search digest is often too thin to support a responsible answer.

The budget is 10 steps and a 150-second wall clock. A timeout is not a wasted run: the researcher writes as it works, so the runner returns the accumulated partial instead of coming back empty-handed.

An answer with a confidence rating

The report carries more than a conclusion:

Named sources — every claim hangs on a citation.
An explicit confidence and controversy assessment — what is solid, what is still disputed. "The field has no consensus on this" is part of the answer, not a defect.
Suggested follow-up pages — where to start if you want to dig yourself.

What you get isn't just an answer; it's a researcher's professional judgment: what the answer is, how sure they are, and where people disagree.

It gets better at how you ask

The researcher keeps its own playbook (memory/agent-playbooks/search.md) and its own fitness bucket. Good and bad research gets scored, and enough score triggers an evolution of that playbook (see The Evolution Loop). Over time it gets better at researching the way you ask questions.

Setup

DEEPSEEK_API_KEY is required. The web-search infrastructure call always goes through DeepSeek, regardless of which model you chat with — the same in both forms (Vercel deployment / local npm). Without that key, webSearch fails with an explicit error instead of silently degrading into a guess — under a pure-bridge (zero-key) local configuration, it is simply unavailable.

Recall — ask recall about the past; ask the researcher about the world
Colleagues, Not Tools — why the main agent outsources research to a colleague
The Evolution Loop — how the fitness score drives playbook evolution

A researcher, not a search box

How it researches

Two steps, both in its own hands:

The search itself is provider-executed, producing the candidate material.

It reads pages itself. Its own webFetch opens the most promising pages for a real read — at most 3 pages, quota-bounded — because a search digest is often too thin to support a responsible answer.

An answer with a confidence rating

The report carries more than a conclusion:

Named sources — every claim hangs on a citation.

An explicit confidence and controversy assessment — what is solid, what is still disputed. "The field has no consensus on this" is part of the answer, not a defect.

Suggested follow-up pages — where to start if you want to dig yourself.

What you get isn't just an answer; it's a researcher's professional judgment: what the answer is, how sure they are, and where people disagree.

Setup