Each row: id (the company id query_companies uses, and the company_profile_id on query_people and query_search_results rows), display_name, domain, spine (industry, location, employee_count, description and the other firmographics), and:

• people_count and people_preview: the people the user knows at the company, across the account, not only the ones an agent found. To list them with their outreach, call list_people with company_id and no agent_id.

• signals: {count, latest, monitored}: how many signals the company has, the newest one, and whether a monitor watches it. With agent_id, these are this agent's (from its searches and its monitoring); without it, every agent's search finds in the account plus the user's own monitors.

• signal_entries, only with include_signals: every one of those signals, newest first.

With agent_id, each row also has:

• lists and sources: the search lists and discovery tools that surfaced the company, empty for one the agent only monitors. columns holds the researched [label, value] pairs.

• verdict, verdict_reason and rejected_on: the curation recorded on the company's search rows, as the Verdict column shows it. 'rejected' when every row is rejected, 'qualified' when any row is qualified, else 'unclear' (screened but too thin to judge) when any row is, else null, with the reason from the latest row carrying it. On a group_by='list' bucket page, or under a "run" filter whose runs all saved to one list, it's the verdict on that list's row. rejected_on is {lists, reason}: the lists whose own verdict is 'rejected' while the company's verdict isn't, with the latest rejected row's reason; its lists is empty otherwise and on those two reads. A company rejected on every row that the agent doesn't monitor is left out unless a "verdict" filter asks for 'rejected', a "run" filter is set or include_rejected is set, the same as the user's default view.

• criteria, only with include_criteria: why the agent matched the company, as its opened row shows it. Per list it's in, {list_name, evaluations}; each evaluation is a criterion with the reasoning, a verdict (satisfied 'yes', 'no' or 'unclear', or a numeric score where 7 and up is met) and http(s) references.

Criteria run a few KB per company and signal entries usually under one, so ask for either on a narrow read (a q, one bucket, or a small limit) or from run_code.

The "list" and "source" groupings, the "list", "source", "verdict", "run" and researched-column filters, sort='latest_find', include_criteria and include_rejected read one agent, so they need agent_id.

To count companies per group, pass group_by alone. To list one group's companies, pass group_by plus a bucket key taken from those counts. Otherwise you get every company, in the order sort picks. Page until next_cursor is null by passing it back as cursor. A direct call returns 25 rows by default; for a wide pull, call this from run_code with limit up to 200 (nothing truncates there) and print only the fields you need. A dict. A group_by without a bucket returns group_counts, a list of {key, count} with the largest group first and the '' group last, and total, the matching companies each counted once. With group_by="list", each group also carries criteria: the search criteria a Websets list screens its companies against, '' for any other list. Every other call returns results (rows shaped as above) and next_cursor, plus total (every matching company) when group_by is omitted. With agent_id, every call also returns filtered_out_count: how many companies this same call matches once a "verdict" filter on 'rejected' is added to filters, with the same q and group_by and no bucket, each company counted once. Only under the same search, filters and grouping as the user's Companies table does it equal the number that table shows as "didn't match".