Interceptors and demons: two ways your agent joins the conversation

Personal AI agents are about to be stan­dard issue — one or more per employee, the way lap­tops were in 2005. The ques­tion that inter­ests me isn’t whether they show up. It’s how they show up when two humans are talk­ing to each other.

Two struc­turally dif­fer­ent pat­terns are in play. They look sim­i­lar from the out­side — both pro­duce mes­sages enriched by an agent — but they have oppo­site con­se­quences for trust, attri­bu­tion, and how anyone can argue with the agent. Naming them apart is the first step to think­ing clearly about which one your team is actu­ally adopting.

Pattern A: the interceptor

Team lead speaks unfiltered. They send a voice-noted frag­ment, raw, half-finished sentences. The agent inter­cepts, attaches the project con­text, runs a quick “you said the oppo­site on Tuesday” check, and ships a pol­ished mes­sage under the human’s name. The recip­i­ent’s agent does the reverse: parses the inbound, pulls related arti­facts, decides whether to escalate.

On the sur­face, one mes­sage per human turn. The agent is invisible.

Teams already run this pat­tern in narrow form — AI drafts emails for humans to send, but rarely sends them itself. The line between draft-assist and ship-without-asking is the only thing keep­ing the org-chart fic­tion intact: humans are still the only account­able actors — for­mally, the human said it, signed it, owns the consequences.

The price is that every­thing the agent does is unauditable. You can’t see what was added, sup­pressed, or pulled but left out. And there’s no honest way to dis­agree with your own agent in public — if the con­tra­dic­tion note it inserted was wrong, cor­rect­ing it means either look­ing like you con­tra­dicted your­self or admit­ting an agent ghost­wrote the message.

Pattern B: the live participant

Each human has their agent in the con­ver­sa­tion as a named participant. I like to call them demons in the clas­si­cal sense — enti­ties that shadow their human, read every­thing the human reads, speak on their own line. A two-person DM becomes a four-actor conversation.

Every con­tri­bu­tion is attrib­ut­able and overridable. You can dis­agree with your own demon pub­licly — “no, that con­tra­dic­tion note is wrong because…” — which becomes a learn­ing signal the agent can train on.

The down­side is loud. Density goes up two to four times. The “always inter­ject” fail­ure mode is real and most early imple­men­ta­tions will live in it for a while. Norms about when a demon speaks versus stays quiet don’t exist yet and will have to be invented per team.

The cultural delta, not the technical one

Interceptor is harder to imple­ment but easier to adopt. Live-participant is tech­ni­cally triv­ial and cul­tur­ally enor­mous, because it makes the agent’s con­tri­bu­tion leg­i­ble and there­fore contestable.

I expect most orgs to start with inter­cep­tor and qui­etly migrate to live-participant. Attribution is one reason — once the agent shows up on its own line, you can cor­rect it, over­ride it, and see what it actu­ally did. The bigger reason may be con­text: a named agent in the thread is one you can hand fresh con­text to mid-conversation, instead of hoping the har­ness loaded the right files. Attribution makes it safe; context-supply makes it useful.

Mixed-by-surface probably wins

Live-participant fits chat-shaped surfaces. Slack, Telegram, group rooms — attri­bu­tion is already cheap, den­sity is already high, one more line per actor doesn’t break anything.

Interceptor fits artifact-shaped surfaces. The Linear issue body, the email, the PR description. The unit is the arti­fact, not the turn. The agent’s job there is to pro­duce one clean output, not to talk about it.

The unsolved prob­lem is the handoff. When a Slack thread decides to file a Linear issue, the same agents have to shift mode. Whose demon writes the first ver­sion? Does the other get to review it before filing, or only after? Nobody has a clean pro­to­col yet, and who­ever ships one will get more credit than they expect.

The interface barely exists yet

The inter­cep­tor pat­tern only has one mature UI today — email auto­com­plete and AI draft-replies. Live-participant has Slack and the chat surfaces. Everything else is still raw, and the next sur­face to mature is almost cer­tainly voice. People are dic­tat­ing more, voic­ing notes more, talk­ing to their phones more, and the gap between what you say and what the recip­i­ent should read is exactly where an inter­cep­tor wants to live. I’d expect the first seri­ous voice-interception UI within a year or two. Whoever ships it well will set the con­ven­tions for every­one else.