Introducing CodeGrid! The terminal workspace for AI coding agents. Run dozens of Claude Code another agent sessions in multiple infinite canvases. Drag, resize, and zoom each pane. Status indicators show what every agent is doing, across all workspaces. Native macOS. ~10MB. Open source. codegrid.app

hot take: 'multi-agent' doesn't mean six agents running. it means six agents talking. parallel isn't a team. communication is a team.

week 1 of posting daily about CodeGrid done. wrote 35 posts, answered every reply. the obvious early signal: the attention system resonates way more than I expected, and the canvas less. people don't know they have the spatial problem until they feel it. noted.

if you could only keep ONE: · multi-agent canvas · agents talking to each other · native notifications when agents need you · git UI next to the agents

broadcast also works as a poor man's benchmark: same prompt, six agents, watch who finishes first and who writes better code. you'll develop opinions fast.

⌘B turns one keyboard into six. broadcast mode, plus the honest recap of week 1: • broadcast: every keystroke goes to every pane. one prompt, six agents answering in parallel. • use it to ask the same question to Claude, Codex and Gemini and compare the answers side by side. • use it for fleet hygiene: /clear everywhere, git status everywhere, the same setup command everywhere. • recap: this week we covered the canvas, native speed (~10 MB), local-first (0 telemetry, MIT), workspaces, and attention. • next week: the agents themselves. all six. and the thing that makes them a team instead of six tabs. what should I go deeper on? replies steer next week's threads. genuinely.

the most underrated multi-agent feature is the dumbest one: type once, send to everyone. orchestration starts with a megaphone, not a framework.

stages of multi-agent grief: 1) denial ('it's still thinking') 2) anger (it was waiting on Enter) 3) bargaining ('I'll check every 5 min') 4) depression (tab #9) 5) acceptance (get a tool that taps your shoulder)

what's the longest an AI agent has sat waiting on your yes/no while you were in a meeting? I'll start: 1 hour 4 minutes.

live activity indicators are readable even zoomed all the way out: idle, running, waiting, error, dead. five states, five glyphs, one glance.

your agent finished 20 minutes ago and you didn't notice. CodeGrid's whole attention system, explained: • it watches every pane for the states that matter: Y/N prompts, approvals, errors, 'waiting for input', done. • a global attention bar shows fleet status at a glance, even with 9 panes running. • ⌘⇧A jumps to the next agent that needs a human. inbox zero, but for your fleet. • native macOS notifications fire when an agent finishes, errors, or asks. the dock badge counts what's waiting. • menu-bar status means you can be in another app entirely and still know. agents are interns with infinite stamina and zero initiative to interrupt you. the tool has to do the interrupting.

AI agents don't fail loudly. they fail politely, sitting there waiting for a yes. the real innovation isn't smarter agents. it's knowing the moment one needs you.

feature we killed: scanning $HOME to auto-discover projects. felt smart, was creepy and slow. replaced it with a plain most-recently-used list. lesson: the obvious boring solution usually respects the user more.

how many projects do you actively touch in a normal week? · 1 · 2-3 · 4-6 · I've lost count