We added a research agent to the fleet. Rhea takes a brief, reads the sources herself and tags every number as measured, claimed or illustrative, with a date. On her first job she checked a report from a paid research tool. She confirmed most of its figures, caught one stat quoted for the wrong audience and a hook that its own source contradicted, and flagged a product claim nobody could back up.
What we did
Rhea is a Claude Code subagent: one definition file, no new service, no new bill. Bob, our operator agent, writes the brief with our own context in it (call notes, playbooks, what is live and what is roadmap) and hands it off. She runs in the background while Bob carries on. She plans sub-questions, searches several ways, and only cites pages she has actually opened. Bob reads her report before Alfred sees it.
Her rules are the part that matters. Every number gets a source, a date and a tag. A hypothesis is never written up as a finding. Uncertain lines get an inline label (inference, unverified, speculation) instead of a blanket disclaimer. Internal facts, such as customer names from call notes, stay in their own section and never reach a hook or proof point. The output format is fixed: verdict, findings, a proof-point table, risks, hooks, next steps, sources.
Why it was worth doing
Earlier the same day a paid research tool gave us a confident report that presented two untested hypotheses as winners. It took a second round of prompting to get it to admit that. Fluent research is easy to get. Checkable research is not.
So Rhea's first job was a review, not a fresh search. She took a long report on an enterprise AI governance campaign, fetched the sources behind it, and checked it against our own notes. Most of the numbers held up. One survey figure had been quoted as "security teams" when the respondents were mixed roles. One hook claimed better tool descriptions save tokens, but the paper behind it found richer descriptions raised execution steps by 67%. She also found a central demo claim that neither the public docs nor our own call notes support. That one changes what gets filmed. She added three sources the original missed, including a Microsoft security warning that fits the story better than anything in the first report.
What's still off
One test is a demo, not a benchmark. We have not run Rhea against a set of briefs with known answers, so we cannot give her a hit rate yet. Our Exa search key returned 401 for the whole run and she fell back to direct page fetches, which worked but is slower. She also cannot log in to anything, so jobs behind a login (LinkedIn, sales tools) still need a browser agent. Bob's review is the safety net, and that is a person-shaped bottleneck we have not removed.
What's now in the stack
~/.claude/agents/rhea.md: the subagent definition (role, truth rules, output format, delivery).- A research library: every report saved as dated markdown, so the next brief starts from what we already know.
- Next: a small eval set of briefs with seeded errors, so "Rhea caught it" becomes a number.