An agent can host several named monitors. mode='network' watches the user's whole 1st-degree feed; mode='list' watches an explicit set of people and company pages — pass members as profile URLs or vanity slugs for people, and the full linkedin.com/company/... URL for a company page (a bare company name or slug is read as a person's slug); mode='topic' searches all of LinkedIn for posts matching keywords, beyond the user's network. Watching the user's own profile (their URL in members) is how a user gets engager capture on their own posts. writing_instructions steer the comment voice. An edit to an already-active list monitor waits for its next scheduled run, because every list run bills each watched profile; every other call kicks off a run immediately, so a new monitor or an edit shows results within minutes. After that it runs on run_weekdays (default Mon/Thu, in the user's timezone). The first run covers only the members present at creation, and edit_monitor_members starts no run — so create a list monitor with its complete membership in this one call (assemble a large or query-derived list in run_code and pass the whole variable), since members added afterward miss the backlog their first scheduled run would have caught. Comments land in the approval queue by default — they post without review only if the user has enabled auto-approve for the linkedin_comment subtype.

list mode works even when the user hasn't connected LinkedIn: discovery is cookieless and drafting never touches their session, so drafted comments hold in the approval queue and post once they connect. network/topic discover through the user's own LinkedIn session, so those two need it connected first.

Schedule and recency are per-monitor. run_weekdays is which days the monitor runs — weekday ints Mon=0 … Sun=6 in the user's local timezone (e.g. [3] = Thursdays only, [0,1,2,3,4] = every weekday). recency_days is how fresh a post must be to earn a drafted comment: only posts published within that many days are commented on (default 7, max 7). Set these when the user asks for a cadence ("once a week on Thursdays") or freshness rule ("only comment on posts from the last 3 days") — those are the knobs, not the agent goal text. A recency_days shorter than the gap between runs deliberately skips the older posts in that gap: narrow it for a user who wants only the freshest posts, or to trim a list monitor's per-post cost (see Cost — a shorter window pulls fewer posts, though the per-profile floor dominates a large list's cost). Leave it at the default week to comment on everything since the last run. Omitting run_weekdays, recency_days, or (on a list monitor) members on a re-setup keeps the monitor's current value; the other fields overwrite on every re-setup, so re-pass writing_instructions/draft_comments/fetch_engagers/capture_authors/ engager_icp/post_filter/engager_instructions when editing rather than dropping them. For membership tweaks (add or drop a few people) use edit_monitor_members — it takes only the changes, so you never re-echo the whole list (dropping a URL on the re-echo silently stops watching that person).

Cost: a queued drafted comment costs 1 credit (all modes). list mode additionally bills for the paid cookieless scraper's per-event usage (1 credit = $0.10): 0.04 credit per post it returns plus 0.02 credit per watched profile that posted nothing in the window — every watched profile is billed on every scheduled run whether or not it posted, so the per-profile floor (not the per-post charge) dominates a large list. That floor is a hard minimum: a run costs at least 0.02 × number of watched profiles (a silent profile exactly that, a posting one more), so a 2,000-profile list is ≥ 40 credits every scheduled day. Because that recurs per run, recommend a reduced cadence for a large list: the default is Mon/Thu, and for lists in the high hundreds or more, suggest weekly (run_weekdays=[0]). Other levers are trimming the member list; narrowing recency_days only trims the smaller per-post component. Flag this to the user when setting up a list monitor, especially a big one; a zero-balance list monitor won't run at all (even for engager capture). network/topic discovery is free (Unipile session).

The capture actions share two optional companions, both handed to the woken new_post_engagers run on the event. engager_icp filters WHO is collected, engagers and authors alike (ICP, titles, exclusions); blank collects everyone. engager_instructions decides WHAT HAPPENS to them; blank means they are recorded into the agent's Output list and nothing else is done. Capture never sends outreach on its own — if the user wants a sequence or a connection request, it has to be spelled out in engager_instructions.

keywords is a LinkedIn boolean query. Combine topics with OR and quote multi-word phrases: "creator economy" OR "influencer marketing". AND narrows, NOT excludes, and parentheses group: ("seed" OR "series a") AND fundraising. Operators must be UPPERCASE (lowercase and/or/not are read as literal words). The combined AND+OR count is capped by the user's LinkedIn plan (free/premium 5, Sales Navigator 15, Recruiter unlimited; NOT doesn't count) — past the cap LinkedIn silently returns nothing, so cover many topics with several monitors each within the cap, not one giant OR. A single unquoted topic (fundraising) is fine. A malformed query (unbalanced quotes/parens, dangling operators) is rejected.

For network, leave keywords empty in almost all cases — the feed is already ranked and the per-post draft step skips anything not worth commenting on; set a narrow query only when the user wants one specific slice. list mode ignores keywords — it watches the named people's and pages' recent posts directly, and the draft step is the relevance gate. For mode='topic', keywords is required and is the only filter.

Reusing an existing name updates that monitor (and re-activates it if it was stopped) — this is also how you edit one. Prefer edit_monitor_members for a list monitor's membership; reserve setup's members for the monitor's full initial membership or a deliberate full reset. Dict with status ('active' or 'error'), name, message, and additional keys.