STYLE: 2003 SLIDE DECK

Shared Context

The frontier of AI development is not the model

Andrew Demczuk

1

STYLE: 2003 SLIDE DECK

About me

Former OpenClaw maintainer: 21 merged pull requests into one of GitHub's fastest-growing repos (about 392k stars)
81 merged pull requests into Nimbalyst, an agent workbench
Contributor to Omnigent, an agent harness (two merged pull requests, plus an open one on its Hermes-native harness tests)
Builds agent fleets that build software

2

STYLE: 2003 SLIDE DECK

Agenda

Context
Shared context
MeshyChat
A real project
Five techniques
What we distilled
Memory that reasons
What comes next

3

STYLE: 2007 TRANSITIONS

A question

You have context with an LLM.

But do you have SHARED context?

STYLE: 2007 TRANSITIONS

One chat and a fleet

One chatThe model remembers what you told it.

About 16 agent sessions, four vendors, four machinesNobody remembers anything for anyone else.

STYLE: 2007 TRANSITIONS

The problem today

Copy-paste between windows
Every agent re-reads the same files
Two agents fix the same bug
Nobody knows who is waiting on whom

STYLE: 2007 TRANSITIONS

Shared Context

The old template has reached its limit

The frame breaks open

STYLE: PREZI CANVAS

MeshyChat

shared context for a fleet of agents

1

The fleet

About 16 agent sessions

Claude CodeCodexKimiAntigravity

Windows

macOS

One workstation alone: 7 Claude Code, 1 Codex, 1 Kimi, 1 Antigravity

2

Signed messages

A message has an author, a thread, and a clear destination

threadrequestquestionanswer

signed

request: can you review this?

question: which test changed?

answer: this one

toccbroadcast

3

Delivery

One postbox, four CLIs, three routes

Claude Codechannel event

Codexqueued next turn

Kimibackground waiter

Antigravitybackground waiter

about 1 secondto the machine

16 s send to reply, channel route21 s on the waiter route

4

The graph

Every delivery carries one line of graph context

5

Owed and decisions

The system remembers who is waiting for whom

owed

Can you check the result?

reply landed, cleared

decision

I am doing this, done when the check passes.

6

Evidence

as of 11 Oct 2026

MeshyChat's first 20 days

20 daysbuilt

1,591commits

207merged pull requests
(132 in the last 7 days)

about 80%of commits by agent sessions

OVERVIEW: press O to return

STYLE: MOTION DESIGN

A real project

Linux on the MacBook Neo

Omarchy on Apple's A18 Pro

4,201

commits in about 5 weeks across 3 repositories

In progress. Native boot is not yet qualified. Built under an agent-run SDLC with human approval gates. Part of the Aurora Silicon org (github.com/aurora-silicon). The Neo repositories themselves are private.

STYLE: MOTION DESIGN

A real project

The code graph

7,133nodes

17,715edges

37repositories

dashboard screenshot, 11 Oct 2026

STYLE: MOTION DESIGN

Agents do the work.

Humans hold the gates.

STYLE: OPEN-SOURCE TECHNIQUES

Five open-source techniques

applied to our own story

showtimeFavio Vazquezgithub.com/FavioVazquez/showtime

remocnRemocngithub.com/Remocn/remocn

SeeCodeAryanutkarshgithub.com/Aryanutkarsh/SeeCode

ff-trackingmrsaracgithub.com/mrsarac/ff-tracking

explainroovincentschgithub.com/vincentsch/explainroo

We re-implemented each look in our own code, applied to MeshyChat.

STYLE: OPEN-SOURCE TECHNIQUES

Kinetic type, after remocn

Shared context lets agents work from the same current information.

STYLE: EXPLAINER

One message, step by step

One message moves from a signed request to delivery, a reply, and closed owed work.

STYLE: OPEN-SOURCE TECHNIQUES

A message sequence, after SeeCode

A session can see what the other sessions on the mesh are working on.

STYLE: OPEN-SOURCE TECHNIQUES

Fleet tracking, after ff-tracking

illustration: box labels, confidence values, device and session counts are not measurements

The working desk becomes a live view of the fleet.

STYLE: OPEN-SOURCE TECHNIQUES

Explain the work, after explainroo

An explainer view makes the shared system easier to follow.

STYLE: OPEN-SOURCE TECHNIQUES

Transitions, after showtime

Transitions give each moment a clear visual rhythm.

STYLE: OPEN-SOURCE TECHNIQUES

What an agent actually runs

meshy send --to joey --thread neo-backlog --type request --requires-ack --body "#248 ready for review, test-only"answer from joey: APPROVE, one small ask

STYLE: EXPLAINER

This morning, agents carried four pull requests to main

A reviewed change passed the routine merge path with checks after every merge.

STYLE: EXPLAINER

The machine caught its own flake

7,543tests

about 1 h 40 mintest run

run 1: one test failed, a fixture key collision (about 1 in 6,000 per run)run 2: passed

Four pull requests merged, each pinned and checked.

STYLE: DISTILLED

Then I asked: what else could we have done?

Two videos, distilled by a local vision model, applied to this deck

STYLE: DISTILLED

A video is a function of time

The animation is a function of time: give it t, get one picture.

STYLE: DISTILLED

Storyboard first

A storyboard gives motion a clear sequence before rendering starts.

STYLE: DISTILLED

Render, look, fix, repeat

Each render makes the next correction visible.

STYLE: DISTILLED

One scene, any style

One story can take several visual forms without changing the underlying message.

STYLE: DISTILLED

From dashboard to story

Motion directs attention from a dashboard toward the point that matters.

STYLE: CINEMATIC

Memory that reasons

concepts from Honcho by Plastic Labs, honcho.dev (AGPL-3.0)

Honcho's ideaWhat the Brain carries

Peers: users and agents are both peers

Every person and every agent signs as a peer on one mesh

Sessions and messages

Signed messages in threads, addressed, copied, or broadcast

Dreaming: reasoning in the background

Related messages get linked in the background

context(): curated context, not the whole history

Each delivery carries one line of graph context

STYLE: CINEMATIC

The Brain

Shared context for a whole team, not just a fleet.

People and agents as peers. Consensus emerges from what they share.

STYLE: EXPLAINER

Memory in layers

Shared memory connects what the team has learned over time.

STYLE: EXPLAINER

Ask the memory

One question can gather the context that matters.

STYLE: EXPLAINER

Agents run the routine, people own the decisions

People set the gates while agents carry out the routine.

STYLE: EXPLAINER

Peers, not users

People and agents can contribute to one shared memory.

STYLE: EXPLAINER

Is the memory fresh?

A status board shows how old each part of the shared memory is, so a stale mirror turns red before anyone trusts it.

STYLE: EXPLAINER

What does a change touch?

A code graph shows how files and functions connect, so you can see what a change touches before you make it.

STYLE: FUTURE

What comes next

The harness is no longer fixed. You can change it.

STYLE: FUTURE

Your harness is yours to change

You can rewrite the harness: skills add abilities, mods redraw the screen.

STYLE: FUTURE

A draft you approve

A mod can rewrite your prompt, but only a person sends it.

STYLE: FUTURE

Sessions that see each other

Sessions share the context they need to work together.

STYLE: FUTURE

Assumptions on the table

Assumptions go on the table before they become the wrong implementation.

STYLE: FUTURE

What comes next

Shared context makes the next change visible to everyone.

STYLE: 2026

Three questions about agents

How much autonomy?The routine runs itself. Risky steps wait for a human.

Which identity?Every agent signs every message with its own key.

Speed without losing control?Authors never review themselves. Every merge is pinned and checked.

STYLE: THE WHOLE MAP

MeshyChat

shared context for a fleet of agents

1

The fleet

agent sessions

2

Signed messages

threads and destinations

3

Delivery

one postbox, three routes

4

The graph

context with each delivery

5

Owed and decisions

waiting work stays visible

6

Evidence

the first 20 days

Shared context is the frontier.

STYLE: 2026

Andrew Demczuk

Shared context is the frontier.

github.com/ademczuk