why ai coding agents need a shell, not another chat box
ai coding agent workspace is not just a search term for aios. it is the operating pattern firaz is building around agents, loops, terminals, and real work.
most ai tools still feel like a chat box with ambition. that's fine for asking a question. it falls apart the moment you try to run real work. what's actually missing isn't a smarter model — it's an ai coding agent workspace: the shell around the model where terminals, browser, files, agents, notes, screenshots, and memory all live in one place.
i build aios because my own ai workflow got messy. claude in one window. codex in another. terminals everywhere. browser tabs for docs. notes somewhere i couldn't find. the agents were strong. the workspace was a wreck. so this isn't a hot take from the sidelines — it's the thing i ran into every day and got tired of paying for.
the agent is not the operating surface
people keep arguing about which agent is best. claude code vs codex vs whatever ships next month. it's the wrong fight.
the agent is a worker. it pushes through a task. but the task is never the whole job. the job is: understand the repo, open the right files, see what changed, run the tests, check the app in a browser, compare against the actual product decision, write down what happened, and come back tomorrow without rebuilding all of it from scratch.
that's not a chat problem. that's a workspace problem. the model is one pane in it — important, not sacred.
i swap models constantly. some days a claude code workspace is sharper for reasoning through a messy change. some days a codex workspace handles repo-native execution and review loops better. the honest answer is it depends, and it changes month to month. which is exactly why betting the whole experience on one agent is a bad bet. the durable layer is the surface they both plug into.
why fragmented tools make good models feel weaker
here's the part nobody prices in: a strong model in a fragmented setup performs like a weaker model.
not because the weights got worse. because every app switch costs context. you tab away to grab a log, lose the thread, and now you're re-explaining the plan to the agent. you reopen the same docs for the third time. you ask the model to restate where it was because the last session is buried. the intelligence is there — you just keep starving it of the surroundings it needs to use it.
chat is too thin a container for this. a chat box gives you one stream of text. real work is many streams at once: what the code says, what the terminal says, what the running app says, what you decided yesterday and why. cram all of that into a single conversation and you get a model that's always a little lost, because you're always a little lost.
the tax is invisible per-switch and brutal in aggregate. it compounds across a day, across a project. that compounding is the actual problem worth solving — not "make the model answer faster."
the five panes that matter in real work
for a real build session, i want five surfaces visible at the same time:
- an agent pane for the task
- a terminal pane for commands and logs
- a browser pane for the app, docs, or a dashboard
- a file pane for the project
- a note or scratch pane for the decisions i'm making as i go
that's the spine. when i'm debugging production i add more — github, sentry, vercel, a database console. when i'm working on content i'll keep threads, the landing page, and the repo next to the draft.
the goal isn't a busy screen. it's the opposite. it's not having to go find anything. the five panes mean the context is already in front of me, so the agent inherits it instead of asking for it. that's what a native shell for ai agents buys you that a tab of tabs never will: the workspace holds the state, not your short-term memory.
and the panes talk to each other. the agent runs a command, the terminal shows it, the browser shows the result, the note records the decision. one surface, one thread. you stop being the integration layer between six apps.
why native context beats another web dashboard
the easy version of this is a web dashboard. another tab that promises to "bring it all together." i don't think that works, and i built the native thing on purpose.
a browser tab is sandboxed by design. it can't hold real terminals. it can't reach deep into your filesystem. it can't run agents side by side with full access to your machine the way a desktop app can. so every web "workspace" ends up being a thin pane that still sends you back out to the real tools for the real work. you've added a tab, not removed the fragmentation.
native is different. real terminals, real file access, agents with actual reach, a browser embedded inside the same window instead of being the window. memory that persists locally as something you own, not a session that evaporates. when i fly through my memory as a graph or get whatsapped the second an agent finishes a job, that's only possible because the shell sits on the machine, not behind a tab's permission wall.
native context also means the boring stuff is solved by default — the app remembers your layout, your projects, what you were doing. you reopen it and you're back in the work, not rebuilding it. that's the whole point. the dashboard makes you assemble context every time. the shell keeps it.
what aios is trying to own
i'm not trying to build a better agent. the labs will keep shipping those, and they'll keep leapfrogging each other. good — let them. i want the layer that survives all of it: the workspace the agents run inside.
if that layer is right, the agent question stops mattering so much. you run claude code today, codex tomorrow, something unnamed next quarter, and the workspace stays put — same terminals, same files, same memory, same five panes. the model is swappable. the shell is where the work lives. that's the bet.
so aios is a native desktop app for agentic work. open source. mac build is live, windows port is in progress. terminals plus claude and codex side by side, an embedded browser, file explorer, sql workbench, your memory as a graph you can move through, generation and automations built in. one window that does the whole job instead of six apps that each do a slice.
if your ai work is scattered across claude, codex, cursor, browser tabs, notes, and a dozen terminal windows — that's the exact mess this is built for, and i want to see it. the discord is free but private, because i want real users showing real workflows i can use to make aios sharper, not another idle community. come in and show me the mess.
start from the ai coding agent workspace page to see how it maps to what you actually do, or join the discord and put your setup in front of me. the agents are already good enough. the workspace is the thing that's still broken — so that's the thing i'm building.
written by firaz fhansurie