One map of every AI inside an agent.
Season 2, Episode 4 of "Superpowers" – documenting how we're replacing the media agency for smaller businesses, so they can capture leads they'd otherwise never get. Follow Season 2 here: smoc.ai/blog/categories/superpowers. Season 1 of 5 Minutes is here: smoc.ai/blog/categories/5-minutes-series
Website chat, a card flow, and a LinkedIn agent look like three products. They are three ways through the same system. The chart above is that system. AI does the work in each of the four jobs, and a tag on each piece says where it runs: every agent, the website, or LinkedIn.
Three agents, one graph
Website chat qualifies someone who is already on your site or a landing page, and hands off a lead.
Card flow is that same conversation with one step on the page. The public link ends in /cards. Chat stays in bubbles. Both are web flows, and both use the same Studio graph.
LinkedIn runs outbound: a profile view, a connection request, a message, an InMail. The reply lands in Inbox, and a person still takes the thread.
01 — AI reads the company
AI reads the site before a flow is written. It crawls the pages you point it at — about, product, pricing, who you sell to — and turns them into a commercial profile: the offer, the audience, the voice. Brand comes with that: logo, colors, images.
Chat, cards, and LinkedIn are written from the same profile. A thin profile makes a thin agent. Add the standing pages and crawl again before you regenerate.
02 — AI drafts the agent
AI writes the first version in Sales Launch and Studio, from that profile, the goal, and the ICP. You review it before it goes live.
Flow generation is AI writing a website agent and a LinkedIn flow from the profile, the goal, and the ICP.
Theme is AI putting the brand on chat and on cards: colors, logo, images. LinkedIn messages stay text, so this one is tagged Web.
Translation is AI filling the other languages and leaving the graph where it is. Web flows and message flows both take it.
Prospect list is LinkedIn. AI finds the right prospect list, then refills it from the audience that fits while the flow is running.
03 — AI runs the conversation
The person on the other side sees a message, a card, or a LinkedIn note. AI chooses the path, the time, and who is on the list.
Reinforcement learning. Reinforcement learning on the flow path. On the canvas this node is labeled AI. It tries more than one opening and sends more people down the path that is working, while a small share keeps exploring. Episode 2 is that mechanism. A generated flow opens with one. A branch you want fixed is a decision, with the rule written on the edge. A branch that should learn uses reinforcement learning on the path. It runs on website chat, card flows, and LinkedIn.
Recommendation. Web only. AI scores the answers so far and routes to the outcome that fits. A quiz that ends in three different offers is this, with the weights on the answers.
Right timing. LinkedIn only. AI keeps connection requests, messages, and InMail inside the days and hours you set, and puts more of that week on the scheduled days and times that get more accepts and replies.
Reply reading. LinkedIn only. AI marks the first reply as a live lead when it is a next step or a referral. Inbox is still where someone answers.
04 — AI gets better
Recommendations. AI proposes a variant: more leads, more engagement, clearer language, or a cleaner structure. You apply it. That runs on web flows and on LinkedIn flows.
Category learning. After you review a draft and say how this kind of flow should be built, AI uses that note the next time it writes an agent in the same category.
Proven patterns. AI starts a new draft from an example that already works for that category, then adapts it to your profile. Episode 2 called this generating from the best of the network. The chart is where that sits next to the rest.
What you still do
You mark the agent reviewed. You connect LinkedIn if it sends. You go live. The four jobs run around that review.
The conversation those agents run is also what we send back to Meta, Google, and LinkedIn. That was episode 3.
👉 Follow Season 2 here: smoc.ai/blog/categories/superpowers. Season 1 of 5 Minutes is here: smoc.ai/blog/categories/5-minutes-series
Shipped this month — October 2026 release log.
