Agents

Agents

Personas are for talking and writing; agents are for jobs. An agent has a name, a picture and a job; it works on its own and comes back with reports, and you can talk with it about them in the Chat tab (agents are listed there below the personas). What it finds you then write with a persona: Work on it… (a chat) or Write it with… (the Writing tab). The first one is a story hunter: it searches the web for news on your themes and brings leads: a headline, a line or two on what it is and why it is interesting, and the link, many of them, for you to scan and pick from. Set its sites limit to 2 or more and it writes each story up instead, as a small article from that many different sites, with the facts to check.

Reports shows what the agents brought, newest first; new ones in bold. Filter by status, theme and agent. A report has a title, a teaser, the article, its sources (only pages the agent really read in that run) and a list of facts to check, each with the number of its source.

• ↳ follows up: a lead that continues an earlier report (a reply, a consequence, the same people days later) names it; the agent sees what it reported in the last three weeks. Work on it… and Write it with… then hand over both reports, so the connection comes along.
• Work on it… starts a new chat with a persona you pick (the theme's persona if it has one, else the one you picked last time), Web on, with the report typed into the input: add what you want and send.
• Write it with… makes a new project in the Writing tab with a persona you pick: the report and its sources go on the project's Plans page, which the persona always reads; chapter 1 starts empty.
• Keep / Used / Dismiss: Dismiss takes it off the list (the file stays). Kept, used and dismissed stories go into the agent's next runs as your taste, more or less like them; nothing it reported before is reported again.

Agents is where you set one up:

• Picture: its face, in the lists and the chat (kept at once, no Save needed).
• Voice: optional, how it talks if it has a character (a gruff ship's engineer). Only the way it says things: its rules on facts and sources always come first. Empty = plain and neutral, which is best for hunting.
• Models: the ticked launchers, in order. A run uses the loaded model if it is one of them; with nothing loaded it starts the first one this machine has and stops it again afterwards. When another model is loaded (you are chatting) a scheduled run waits and tries again at the next check; Run now asks whether to stop the other model.
• Schedule: manual, every 6h, daily 07:00, 19:00, weekdays 07:00, mon wed fri 08:30. It only runs when Run on schedule on this machine is ticked: that puts one line in your crontab which runs aicove agent due every 10 minutes. A run missed while the machine was off is made up once.
• Style (Factual is good for reports), Thinking (off makes a hunt for leads much faster), Limits: searches, pages read, tool calls, minutes, stories or leads per run, and sites: 1 = leads, 2 or more = written-up stories from that many sites. Each step of a run may think for about 3000 words at most; a model cut off is told to act.
• Search from: where its searches and page reads go out from. Normally the machine it runs on; or another machine of yours that ssh knows (a Host in ~/.ssh/config) and that has a SearXNG as a user service on 127.0.0.1:8081. Then sites see that machine's address, not yours. Its SearXNG is started for the run if it isn't running, and stopped again afterwards.
• Themes, one per line: Name | what it means to you | persona to work on it with.
• Task: what it should do, in words.
• In the chat it may read files: then, when you talk with it, it may read the files the Settings tab allows (a system helper reads system files this way). Never write; runs never touch files.

Talking with an agent: pick it in the Chat tab under Agents. It knows its job and its newest reports (and what you put in its knowledge/ folder), answers in its voice, and with Web on it can check a report or dig deeper. A full run is still started here, with Run now or Research….

Research… (next to Run now) gives the agent one topic to dig into instead: "cooking the perfect steak", "the history of the bicycle", a name, a question. Old sources are welcome there (archives, research, books, news of any year, fringe sites marked as fringe). It reads page after page and notes each source on a card (what the page says, what kind of source it is), then writes a dossier as its report: what is solid, what is disputed or only claimed and by whom, open questions, 3 to 5 proposed articles with angle, hook and sources, and all the source cards. On a dossier, Write articles… has a persona of your choice write articles from it (headline, teaser, text) in its own voice, facts only from the dossier's sources, with wishes like "in German, about 800 words"; they come under Reports too. Research runs are long: give it more searches, pages and minutes than a news hunt.

A lead may come straight from a search result (the report shows what the result said, or that the agent read the page); its address must be one the search gave or a page it read. A written-up story stands on pages the agent read in that run, from as many different sites as sites says. Neither may be a front page (a front page only helps find articles) or older than 3 days when the page or its address says when it came out. A search for an old year ("October 2024": models like the year they were trained in) is sent back. The agent is told today's and yesterday's date. For finding stories a plain voice works best: a character invents angles and critics; let a persona work on the story afterwards (the theme's persona, Work on it…, Write it with…).

An agent may search the web and read pages, and save its reports in its own folder. Nothing else: no files of yours, no commands. Runs go on when the window is closed; Stop ends one. The last run's log is under the form. Small models are weak at long chains of searching and reading; give them a short, clear task and a few themes.

Agents of another machine (one you have launchers of) show under its name. They run there, on its own schedule; editing, Run now and the reports work from here over ssh. Their reports are fetched when the window opens and on Update from other machines.

Where they are: ~/AIcove/agents/<id>/: agent.txt (the settings, by hand works too), voice.txt, picture.png, chats/ and knowledge/ like a persona's, stories/ (one Markdown file per report), seen.txt (links already reported), runs/ (a log per run).

---

Made from the guide inside AIcove (cove/data/guide.md); it opens with the ? button.