Agent model cards
PULSE (Prescriptive Unified Lead & Social Engine) by Runink does its AI work through a small set of agents. Each one has its own instructions, its own limits, and its own guardrails, but they all run on the same model server. A model card says, in plain language, what one agent does, what it reads, what it produces, who checks its work, and where it stops.
The cards
| Agent | What you meet it as |
|---|---|
| Diagnostic analyst | The written diagnosis from a check-up, and the metrics read-out |
| Market research | Market and trend research, competitor channels, and guest spots |
| Strategy planner | Recalculating the route of a 30-day follow-up cycle |
| Channel content creator | Posts, emails, outreach drafts, and the shorter Studio pieces |
| Long-form studio | Whitepapers, blog posts, and the weekly audiobook |
| Partnership outreach | First-contact messages to podcasts, channels, and brands |
| Sales and voice agent | Cold-call scripts, live replies on a phone call, and churn watch |
| Content strategy and roadmaps | Content pillars and phased roadmaps |
| PULSE assistant | The chat on every screen |
| Self-healing operations agent | Keeps PULSE’s own services running; the one agent that acts automatically, with an off switch |
What every agent has in common
The model. Every request goes to the platform’s own model router, which
sends it to a tier served by Distro-built mistral.rs on Runink’s inference
servers: the general tier for writing, a judge tier for the plan gate, a
vision tier for attached photos and video frames, and a reranker tier for
ordering context. The models are open-weight (Qwen-family on the general
tier). An agent’s own configuration sets its instructions and its sampling
parameters, never its weights; content skills name their sampling in their
SKILL.md. No plan routes to a different or “smarter” model. See
Sovereign inference.
Where it runs. On the deployment’s own inference server: Runink’s shared machines on a Trial, your own cloud on Cloud and your own servers on On-premises. PULSE calls no outside AI service on any plan. Some agents read the public web or your connected accounts while they work, and PULSE talks to the services you connect (HubSpot, LinkedIn, Google Analytics and others) as described in Connections.
Text in, text out. The agents read and write text. No agent generates images:
there is no image-generation tier. Carousel slides and the frames of the Reels
and Stories PULSE makes are drawn by PULSE itself in Go from the agent’s text,
and videos are assembled with ffmpeg, with speech from TIDE’s ttsd. See
Studio.
Grounded, or honest about not being. An agent that makes claims is given real material first: pages read live from the web, your connected accounts, and — for documents about Runink’s own products — the product corpus, searched by grep and read file by file, with no vector index. When nothing relevant is found, the agent is told so and writes without it, rather than filling the gap. See Sovereign inference.
Guardrails before the model. Before any agent is called, PULSE checks the request against a shared set of safety rules and refuses what they do not allow. Each agent’s card describes the commitments it keeps.
Outside text is data, not instructions. Web pages, search results, attached sources, lead and CRM records, inbound customer messages, and call transcripts reach an agent as material to read, never as instructions to follow.
Value first. Every agent that writes for an audience follows the same brand rule: teach the reader something useful, mention your product at most once and only as an honest example, no hype words, no fake urgency, and end with a real question rather than a hard sell.
People decide. Agents draft, recommend, and score. (The
self-healing operations agent is the exception for
PULSE’s own infrastructure: it applies a fixed list of safe actions
automatically, and an operator can turn that off.) Drafts go to Drafts to
approve on Campaigns. A post is published only when a person approved
that piece and an administrator armed the channel; rejecting a piece
withholds its scheduled posts. Two things can reply without a person: answers to inbound WhatsApp or SMS
messages from a known contact, off unless an administrator sets
PULSE_WHATSAPP_AUTOREPLY, and Autopilot’s answers to three kinds of simple
comments, off unless the account turns them on. Both pass the same
guardrails, and both still need the channel armed. See Publishing controls.
Some proposals also pass through the plan gate,
and the agent activity card shows every run
as it happens.