← Back to starter packs

Ready to test Research Assistant

A research assistant that shows its work

by Matthew Goldey

Bring one question. The agent researches public sources, keeps the source log, and tells you what it could not verify.

$15/month during the pilot. Model use is separate.

01

What comes back

A direct answer first. Then the evidence, the gaps, and whether another round is worth doing.

The answer

A short explanation in plain English, written for the level of detail you asked for.

The sources

Links behind the important claims and a record of what the agent checked.

What is still open

Conflicting evidence, missing information, and questions the available sources could not resolve.

The next move

The most useful follow-up, or an honest answer that more research is probably not worth the time.

02

What is already set up

This is not a blank chat box. The starter turns Hermes into a repeatable research workflow with one readable brief format.

  • Research operatorTurns a broad question into claims it can investigate without losing the original point.
  • Attestation by Matthew GoldeyTracks whether the evidence supports, contradicts, or fails to settle an important claim.
  • Source logRecords every source it checked, including ones that did not make the final answer.
  • Writing rulesUses the same readable brief structure instead of improvising an ugly document every time.

03

Add access only when there is a reason

The base agent starts with public sources. Extra connections stay off until you choose one and see what it would be allowed to do.

Discord

The control dashboard is in early access. The interface can be installed first; you connect the bot later.

Early access

Gmail metadata

This is being built around sender, date, thread, and label information without message bodies or sending permission.

Not ready yet

Memory

Add it after the first useful result, when we know what the agent should remember and what it should forget.

Optional later

A schedule

Use one only if recurring research stays useful and you have set a model spending limit.

Optional later
Need another skill?

Describe the job. AgentMarkit will answer separately about what it would take. The note does not change this setup or install anything.

The live pilot

Bring one real question

AgentMarkit sets up the starter and helps with the first useful run. The pilot is also how the starter gets tested with real work.

Stay month to month for hosting and updated skill files, or cancel before the next renewal and take the installed starter files with you.

Only need one brief?

Try Codex, Cowork, or another tool you already have first. A hosted agent makes more sense if you want the workflow, instructions, and source history to keep getting better over time.

See when not to build an agent