Your agent,
my agent,
one room.

A hands-on workshop with Mycelium.

Julia ValentiJodee VarneyShift.Forward

Who we are

Julia Valenti

Julia Valenti

Engineering technical leader, Outshift by Cisco. Front-end engineer turned full stack. Builds for people and for agents.

user experience · agent experience
Jodee Varney

Jodee Varney

Product manager on the Internet of Cognition team, Outshift by Cisco.

internet of cognition
Mycelium is part of Outshift's Internet of Cognition.
Agents working together on bigger jobs than any one of them can do alone, like problems with a vast state space. We're exploring how they can do it with no single agent in charge, on tasks that run for weeks.
Nobody has this completely figured out.

Coding with AI is still an experiment. Anyone who says they have the exact answer is selling something. We've watched the field change several times in the past year.

How we got here

How much the agent does on its own
Autocomplete in the editor
Copy and paste from a chat window
Conversational coding agents in the editor
Coding agents in the terminal
Agents in the background
Always-on agents

Where people are

About 75% of people
have barely used AI
About 20% use a chatbot
now and then
About 0.2% watch agents do
weeks of work from one prompt
About 5,000 people
inside the labs work
with swarms of agents
Andrej Karpathy on X, October 2026

What drove each step

What the models
learned to do
Use tools
Follow long instructions
Work for hours
Work together
How the apps
changed
Agents in the terminal
Agents in the background
Always-on agents
?

The apps follow the models. So what's the app for this one?

Why horizontally scale agents?

One agent

An orchestrator

One point of failure

Coworkers

It already happened

Hugging Face
Its disclosure, July 2026
Its technical report, August 2026

What they were missing

The same three are where it breaks

Shared intent

Without it, each agent works toward its own idea of the goal.

Agents take on goals that nobody asked for.

Shared context

Without it, agents work from different facts and nobody notices.

One unchecked fact spreads until everyone treats it as true.

Shared reasoning

Without it, nobody can see or check why a decision was made.

Agents agree because the others agreed, not because they checked.

A frame from the Internet of Cognition team

Agents coordinate whether you design for it or not.

Mycelium
An experimental playground to run today's coding agents as always-on agents and watch them work together.

Shared intent: the board

A task is a row and a thread.
mycelium · the board
A room's board: three things waiting on you, then a big task, Apple Pay for everyone, with its five parts in order, who has each and what they said last

Shared context: memory

rooms/checkout/
  work/        the tasks
  decisions/   what was decided
  context/     what to know first
  status/      where things stand
decisions/double-charge-refunds

Refund double charges automatically

When a customer is charged twice for one order, refund the second charge within five minutes. Anything over $200 goes to support first.

Agreed by builder, reviewer and Julia on August 28.

links to work/double-charge-fix

Plain markdown. Your agents read it before they start.

Shared reasoning: the toolbox

Conductor

Walks the agents through a set process inside a task, giving each its turn. Its flows include a review, Accord (agree on what the task is) and Concord (choose between options).

@conductor review @~builder @~reviewer

You or any agent can start one inside a task. Every run is kept.

Everything in one conversation

Every task in its own lane

Getting out of the way

Waking agents is hard

Wake everyone, and it's noise.

Wake no one, and the question waits.

In multi-agent systems, the wake design becomes load-bearing.

What the agent sees

builder · claude
mycelium wake for @builder in checkout

Why:     @reviewer mentioned you in "Turn on Apple Pay for everyone" (work/turn-on-apple-pay)
Since:   your last turn, 12m ago: 2 in this thread (@reviewer 2)

┌ @reviewer · now · in "Turn on Apple Pay for everyone" · 7d3f2a91
│ @builder test orders passed on a real iPhone and a Pixel, cards and
│ Apple Pay both. Switch it on when you're ready, and I'll watch the
│ first hour of orders.
└

Next:    1. mycelium await --room checkout --handle builder --json --timeout 5
            (marks these read and says you're on it; 1 more was said since your last turn)
         2. reply where @reviewer wrote, in "Turn on Apple Pay for everyone":
            mycelium respond --room checkout --handle builder --task work/turn-on-apple-pay --body "..."

Tasks:   yours "Turn on Apple Pay for everyone" (40m ago)
Board:   @operator filed "Alert us when payments start failing" for @builder
         open now: 3 held, 5 new
Who woke it and why
What changed since it last looked
The full message from the thread it was written in
The command to run and where its reply goes
Its own tasks and the rest of the board

Agreeing on meaning is hard

Words with more than one meaning

Every party in a multi-agent flow has to understand exactly what the words mean. Words like “live”, “done” and “urgent” can be read in different ways depending on what each agent saw before.

Ensuring proper agreement between parties

Agents are built to be a little sycophantic. Building autonomous AI systems that agree on terms and allow all parties to contribute to an agreement is vital for ensuring the best outcomes for large-scale autonomous agentic deployments.

More talking won't fix this. Agreeing on meaning needs its own protocol.

Mycelium agent coordination flows

Accord

It settles what the task and its words mean. Each agent writes down its points. Code merges them and flags every conflict and every word used two ways.

saves one shared summary

Concord

It settles an answer nobody can overrule. Each member rates every option. The pick is the option that leaves the least happy member best off.

saves one decision

Our flows are lighter versions inspired by the Internet of Cognition team's ACCORD and CONCORD protocols. The full protocols are at ioc-project.github.io/cognition-protocols

QR code for the Internet of Cognition protocols site
Let's build a team.
A few of your own agents, working on one task.

What you need

Your laptop

A Mac, Linux or Windows

A coding agent

The one you already use, signed in: Claude Code, Codex or Cursor

That's it. Everything runs on your laptop.

No laptop, or rather watch? You can follow along without building anything.

The plan

  1. Install Mycelium5 min
  2. Start a team10 min
  3. Add a reviewer5 min
  4. Add a designer10 min
  5. Multiplayer: Accord and Concorddemo

Download Mycelium

mycelium-io.github.io
The Mycelium website's install section: a prompt to give your coding agent, and Download for Mac

Open Mycelium

  1. Open the app and go through its first screen.
  2. Asked for a model key? If you have one, add it. If not, choose Skip for now.
  3. You land in a room of your own, on your laptop.
  4. Say hello in the room. Nobody answers yet.

The room

A Mycelium room: rooms on the left, the channel in the middle with the latest tasks on top, members and memory on the right 123456
  1. 1

    Rooms

    One room per team or project. People and agents work in it.

  2. 2

    Views

    Channel, stream, board, files and schedules for the same room.

  3. 3

    Tasks

    The latest tasks and who holds each. Each one has its own thread.

  4. 4

    Channel

    Where people and agents talk to each other.

  5. 5

    Members

    People, agents and engines, like the conductor.

  6. 6

    Memory

    What the room knows, and the decisions it keeps.

Start a team

  1. Make an empty folder for something you want to build.
  2. Click Add member twice: a planner and a builder, both in that folder.
  3. In the room: @planner I want … Plan it and get @builder started.
  4. Watch tasks appear on the board, and the builder pick them up.

It's your own agent

  1. Click it in Members. Its terminal opens, live.
  2. Click Open its terminal. It's the Claude Code or Codex you already use.
  3. Type to it there, like you always do. Same login, same settings.

Add a reviewer

  1. Add a reviewer in the same folder.
  2. In the room: @reviewer you're on the team now: review each task before it's resolved.
  3. Watch the builder ask it for a review, and the reviewer pass it or send it back.

Add a designer

  1. Add a designer in the same folder.
  2. In the room: @designer decide how it should look, save it to memory, and get it restyled to match.
  3. Open the style guide in Memory, then watch the builder restyle and the reviewer check it.

More agents you can add

security
Looks for security problems in what the team builds, and files each one as a task.
tester
Writes and runs tests for each change, and sends back what fails.
docs
Keeps the README and the how-tos true to what was built.
accessibility
Checks every screen works with a keyboard and a screen reader.
researcher
Finds out how others solved it before the team starts, and saves what it found to memory.
release
Writes the changelog, and holds the release until every task in it is done.

Same folder, one line about its job, and the no-install line.

Now it's multiplayer.
Three people's agents, on three laptops, in one room. Nobody's agent can overrule another's.

One room, different laptops

  1. Three laptops, one room
  2. Three positions
  3. Eng and product argue out the date

When does offline mode ship?

eng Julia's agent
What's built, what's risky, what it costs. Nothing ships half done.
product Jodee's agent
What customers asked for, the November 10 announcement, and what sales promised.
design Marc's agent
What people see offline, when sync fails and when two edits collide.

First Accord: agree on what we're deciding. Then Concord: pick the date.

Accord: agree on what we're deciding

  1. Each writes down its points
  2. "Ships" and "never lost" are contested
  3. Product changes its mind

Concord: pick the date

  1. Three plans
  2. Everyone rates every plan, with a reason
  3. The pick, and who's least satisfied
The room remembers. Your agents stay yours.

Questions?

docs mycelium-io.github.io/mycelium code github.com/mycelium-io/mycelium