AgentGridAgentGrid

Build agent teams
that ship

Michael Yagudaev
co-founder, AgentGrid
Pull requests I merged · 2026
154 169 118 201 846 JUNJULAUGSEPOCT ×4.2 Sep → Oct
As a software company you have two choices
Build a software factory
or die.
M. YAGUDAEV
BUILD A
SOFTWARE
FACTORY
OR DIE TRYIN'
AGENT ADVISORY
EXPLICIT TOKENS
A@anon_dev_42· 3h ago
lol another vibe coder who just discovered ChatGPT
▲ 2▼replyshare
Started coding at 12…
to buy more LEGO.
Then I found digital things: infinite LEGO.
The only limits were time and knowledge.
I never wanted to code.
I wanted to build.
Code was always a means to an end.
Michael holding a LEGO Trevi Fountain box
April 2025 · where a lot of you are right now

I started with Claude Code
in a terminal. $ claude

1Unix has a CLI for everythingno API? there's a CLI. Beats driving a GUI.
2Labs train models on the terminalreinforced on tool calls
3Model + harness evolve togetherharness = the agent loop + everything around it
But

The TUI is a terrible
place to spend
your whole day.

A proud 1980s computer nerd giving a thumbs up next to a beige terminal
→ SDKs + desktop apps (Codex, Claude): same local CLIs, no terminal
The terminal's biggest problem

10 terminals. 10 agents.
0 saved sessions.

$ claude_
$ claude_
$ claude_
$ claude_
$ claude_
$ claude_
$ claude_
$ claude_
$ claude_
$ claude_
AgentGrid v1
  • infinite canvas
  • save + reload the layout
  • agents spawn + message peers
  • same session, not a new one
Bottleneck #1 · not tokens
Seeing + steering many agents at once.
agent-gridwebsitemobileresearch
AgentGrid · every agent on one canvas · tabs per project · switch in a clickMUTED · press M
May → June 2026
3× LEVEL 1 · the canvas LEVEL 2 · the factory you can skip straight here
How to start

Talk to one agent.

YOUcontext
AGENTwhat + where to look
WORKERresearch
WORKERresearch
WORKERhow to build it
NEW TEAMthe alternative · still linked
Many agents = focus on one direction, explore the others.
A confession

I reviewed ~1,000
PRs by hand.

Then I noticed I'd stopped
finding anything.

The factory · ticket → merge

Seven stations. No typist.

01
TICKET
the issue is the work order
02
WORKTREE
own tree, own port
03
BUILD
Claude
04
DRAFT PR
real screenshots
05
2× REVIEW
Codex + Grok
06
READY
CI green
07
MERGE
--match-head-commit
■ two ✓ on the same commit = auto-merge
Michael's iPhone screenshot in an iPhone 18 Pro frame
the bug, on my phone
Michael's iPhone screenshot in an iPhone 18 Pro frame
Message Coordinator
dictated it
The dictated message as a sent bubble in the AgentGrid iOS Coordinator chat
sent from the phone
the ticket
the team
draft PR
the PR
the reviewers
review: Claude ✕
review: Codex ✕
re-review: two ✓
merged · 10:38 pm
2h 37m
A day in my factory · two modes
10:30co-founder standup→Granola summary→agents start building
MODE 1 · FIRE-AND-FORGET
  • anything I notice
  • customer requests
  • agents merge + deploy
MODE 2 · EXPLORE TOGETHER
  • UX storyboards
  • research · architecture
  • talked through on a walk
“We have more tickets than agent capacity.”
Mac Mini · 24/7 phone in my pocket |136 merges in our repo on Oct 29
Bottleneck #1: tokens

Tokens were my bottleneck.

19 harnessesswitch harness on the same transcript, manually
2Auto-switch harnesswhen one runs out
3Multiple accountsauto-switch accounts as they run out
4Usage reportspend over time, across accounts, ask the AI about it
How I found it · asked the usage report
michael › Why are my Codex accounts empty in under a day?
⚙ query_usage { group_by: "model", harness: "codex" }
⚙ query_usage { group_by: "mode", harness: "codex" }
Most Codex tokens went to GPT-6.1-Astra (≈5× Sol) with fast mode on (≈2×). Switch the default to Sol and turn fast mode off.
AgentGrid usage report AI panel
The leak I found

3 Codex accounts. Empty in < 1 day.

Astra instead of Sol5×
Fast mode on everything2×
= more tokens than needed10×
→ Sol by default. Fast mode off.
Bottleneck #2: compute

Native apps: heavier builds,
slower incremental compiles.

Now the agents wait on the machine.
AgentGrid Usage history, week of 10/25/2026
The next bottleneck · tokens

Then tokens became my bottleneck.
They aren't anymore.

4.2×
merged PRs
201 → 846 · Sep → Oct
1.2h
median open → merge
was 2.6h
$113→
$48
AI cost per merged PR
list price · Sep → Oct
The dividend: time
Michael's red robot tank next to a VR headset and controller

I finally had time to build this.

The unlock ladder · my blockers
TOKENS ✓unlocked
COMPUTEmy blocker now:
distributed cloud
SELF-EVOLVING
AGENTSthe factory fixes
its own blockers
HARDWAREslow to iterate:
point factories at it
next → Shan Liu, 4:10: “I Built a Self-Improving Agent”
Your factory, tomorrow morning
  1. 0
    Download AgentGrid · agentgrid.sh
  2. 1
    Open your project (or start one), describe the problem
  3. 2
    It files the GitHub issue (automatic)
  4. 3
    It spins up a team: one builder, two reviewers from different model families
  5. 4
    Let it work: gates → commit → merge
  6. 5
    Loop: every 10 min, check for new issues
As a software company you have two choices

Build a software factory.
Or die.

M. YAGUDAEV
BUILD A
SOFTWARE
FACTORY
OR DIE TRYIN'
AGENT ADVISORY
EXPLICIT TOKENS
agentgrid.sh/michael
Start tonight

Download AgentGrid.

This week we onboard
every one of you.

agentgrid.sh/michael
slides, recipe + links:
agentgrid.sh/biossphere
Thank you
QR code: https://agentgrid.sh/michael
@yagudaev
→ ␣ PgDn next build / slide
← PgUp previous
Home / End first / last slide
1-9 ⏎ type a number + Enter to jump
t slide number + timer
r reset timer
n speaker notes
p / 0 demo: play-pause / restart
m video: mute / unmute (starts muted)
v demo slide / slide 15 ask: switch variant A / B
f fullscreen
? this help