Seen from above: streams cross a green marsh, join, and open on a calm bay.

Slack, email and a phone number can all reach one agent through small adapters.

A channel is a door, not the agent. Once the brain lives apart from Slack, email, or a phone number, each new way to reach it is a small adapter rather than a second agent, and the real design questions become where people work and who is speaking.

Three streams cross the marsh and reach the same bay.

Tanush Jagdish · June 2026

An agent’s channel and its brain are separate things. Slack, email, a phone number, WhatsApp, and iMessage are doors; the brain, meaning the files, skills, and memory that make the agent useful, lives elsewhere. Once the two are separated, adding a channel means adding an adapter, not building a second agent. Three consequences follow, and this note covers each: meet people where they already work, keep an explicit trust boundary, and never weld a channel to the agent.

Most people picture a deployed agent as a chat window. When they hear that it can also take text messages, they assume a texting agent and a Slack agent are two products built twice. That assumption is the main thing to unlearn. I put it to clients plainly: Slack or email is just the gateway, and the brain lives in the repository behind it.

#1Start from where people work

Before any architecture, ask one question: where does this person actually work? If someone lives in email and text messages, give the agent its own phone number and email inbox, and make those the entry points. A managing partner lives in email, a founder on the move lives in text messages, and a team lives in Slack. Fit the agent to the person, not the person to the agent. Engineers tend to default to a web app, which is often the wrong door for executives.

What the person gets is an addressable colleague: an agent with its own inbox and, if wanted, its own phone number, WhatsApp account, or Slack presence. Connected to your own inbox, it can do more. Without that, its inbox still sits in its own Google Workspace account, so it can create documents and share them with you from that account, as a new hire would. People email it, text it, or mention it in Slack; nobody has to visit a chatbot.

#2The brain lives elsewhere

Behind the doors, everything is shared: one gateway, one brain, one memory.

One agent, many doorsChannel · gateway · brain
Doors
Slack
Email
SMS · iMessage
WhatsApp
Gateway

Routes, dedupes, manages sessions. Nothing more.

  • · route
  • · dedupe
  • · sessions
One brain

Thinks, acts with its own tools, and writes notes back to memory after every conversation.

one context · one memory

The channel is not the agent. Add a fifth door and the brain behind it never changes.

Many channels in, one gateway that routes messages, removes duplicates, and manages sessions, and one agent with one memory behind it.

Several channels arrive at one gateway whose only jobs are routing, removing duplicate messages, and managing sessions. It hands each message to the agent, which reasons with a cloud model, acts with its own tools, and writes notes back to memory after every conversation. One brain sits behind many doors.

Because the brain is just files in a repository and the channel is a thin layer in front, the answer to “can it also do email and text?” is never “build a second agent.” It is “add a second door.”

#3The build path: email first, then a phone number

Building the pieces in order makes the separation obvious, and it is how I walk clients through it. Start with an agent on your own computer, such as Claude Code, and give it one skill: reading and replying to Gmail. Then, away from your desk, email yourself from your phone and watch the agent on your computer pick up the message, do the work, and reply. Once that loop is clear, moving the agent from your computer to a data center is a small next step.

A phone number is the same pattern with a different adapter. An email channel checks an inbox; a phone channel listens on a webhook, a web address that the messaging provider calls whenever a text arrives. SMS and iMessage work the same way: a webhook adapter at the front, the same brain at the back.

Managed or do-it-yourself, the mechanism is the same; only the operator differs. Do it yourself by putting your documents folder in a GitHub repository and giving the Claude app on your phone access to it, and your brain is reachable from anywhere. Or let a managed service run the same brain for you, so that you can text it a photo and ask it to act. Either way, the brain moves to the cloud and a messaging surface points at it. One client wanted exactly this before a long trip abroad: his brain, on his phone, on the plane.

#4The trust boundary: who am I texting?

A texting agent raises a problem a Slack bot mostly avoids: identity confusion. If an agent answers texts sent to your number, the people texting you cannot tell whether they reached you or software. When it replies without your knowledge, trust erodes quickly: am I texting an AI agent, or am I texting you? That question is the central design decision of the messaging layer, and it has two answers:

  • Control channel: you text your agent to make it do things, and no one else is on the thread. No identity confusion arises. This is the safe default, and what most people want first.
  • Auto-reply persona: the agent answers other people on your behalf. It can be useful, but it needs explicit consent and an unmistakable signal that the other person is talking to an AI.

When I raised this with a client who planned to connect his personal phone, he clarified that he had never wanted auto-replies: the phone was simply how he would reach his own agent. For most people, a control channel is the right place to start. Our own agents follow a related rule: they do not act from a person’s inbox just because a message arrived there, and acting on someone’s behalf requires that person’s explicit opt-in.

Disclosure is also becoming a legal expectation. California’s bot disclosure law, in force since July 2019, bars using an undisclosed bot to mislead people about its artificial identity in order to drive a sale or influence a vote. Transparency obligations under the EU AI Act, applicable since August 2, 2026, require that people be told when they are interacting with an AI system unless that is obvious. Neither replaces judgment, but both point the same way: when an agent speaks to other people, say so.

#5The anti-pattern: channel equals agent

A tempting design welds the channel to the agent: one process, hardwired to Slack. It works until you want email too.

From · the trap

The channel is the agent

Slackits own parallel agent + its own copy of state
Emailits own parallel agent + its own copy of state
Textits own parallel agent + its own copy of state

Three agents pretending to be one, drifting out of sync. Because the process is the agent, the channel can’t move without the agent moving with it.

To · the fix

Adapters share one agent

Slackadapter
Emailadapter
SMS · iMessageadapter
One dispatcher → one agent

A new channel is a small adapter on the dispatcher, not a parallel deployment. Same context, same workflows, same memory.

From “the channel is the agent,” where a new door means a parallel deployment, to channel adapters that share one agent.

Because the process is the agent, the channel cannot move without moving the agent. Email requires a second, parallel deployment with its own copy of the state; text messages require a third. Three agents end up pretending to be one, drifting out of sync.

A fix turns the separation at the start of this note into infrastructure: every channel becomes an adapter on a shared dispatcher in front of one brain. Adding email or text messages means writing a small adapter, not standing up a parallel deployment. Reachable by email and by text, the same agent keeps the same context, workflows, and memory: one colleague with three lines, not three personalities.

#6A round trip in practice

One flow we built for a venture capital firm shows the pattern end to end.

A real round tripThree surfaces · one flow
  1. 01

    Email in

    A partner forwards a deck. The agent reads the email and its attachments.

  2. 02

    Brain

    Writes a structured summary: company, stage, thesis fit.

  3. 03

    WhatsApp out

    Drops the summary on the surface the partner reads first.

  4. 04

    CRM

    Creates the record. The work persists where the team works.

Email in, WhatsApp out, CRM written, and at no point was the channel the agent. The channels were doors; the work was one brain doing one job.

One real flow: a partner forwards a deal email, the agent reads the attachments, a summary arrives on WhatsApp, and a record appears in the CRM.

When a partner forwards a pitch deck by email, the agent reads the email and its attachments, writes a structured summary (company, stage, fit with the fund’s thesis), posts the summary to WhatsApp, and creates the record in the CRM. Three surfaces take part in one flow, with email in, WhatsApp out, and the CRM updated, and at no point is the channel the agent. It is a single deployment, but the shape generalizes.

We have used the same separation internally. Our agent assembled a morning sales brief, covering the day’s external calls and preparation for each, in the brain and delivered it to Slack, because Slack was where we read first.

Naming the agent helps as well. Give it a name, an address at your company’s domain, and a phone number. Each is a labeled door, and behind every one of them is the brain you built.

#7Where this connects

In our curriculum, this note sits in front of Part 4 (the personal brain): the brain you build there is what the messaging layer points at, and its diary and current-state do the remembering whichever door a conversation used. It is the access layer of Part 5 (managed architecture): channel adapters on a dispatcher, and an agent with its own inbox, number, and workspace. It also raises a concern for Part 6 (the institutional brain), where routing each channel to the right person’s brain is the multiplayer problem and choosing between a control channel and auto-reply is a governance decision.

In the research track, Multiplexing: Running Many Agents describes the dispatcher behind the doors, and Brain Design: L1, L2, L3 describes how the caller’s identity decides which brain loads. Olito’s own agents keep channel and brain separate in this way, including OliMed, our agents for independent physician practices, and Agent Oli, an executive chief-of-staff agent available by invitation.

A channel is a door. Build the brain once, and every new way to reach it (email, a phone number, a text thread, a mention in Slack) becomes an adapter, not a second agent. Meet people where they already work, and keep the trust boundary explicit.

More from our research.

Working on the same question?

We look for people to interview, data to compare against, and teams willing to test what we find. Leave a note and we will write back.