What is a coding agent workspace?

Why this category exists

Coding agents got good enough that running one at a time wastes most of your day waiting. Once you run three, the work moves out of the agents and into your head: which terminal has which task, which branch belongs to which chat, and which agent stopped twenty minutes ago to ask you something. A workspace takes that list out of your head and puts it on screen.

Each agent now ships its own way to run several sessions. Claude Code has agent view, the ChatGPT desktop app runs Codex chats in worktrees, and Cursor has its Agents Window. Those are good inside one tool. A workspace is for the moment your work spans several tools, several projects, or both.

What a workspace should do

1. Organize work the way you think about it: projects, then optional sub-projects, then tasks. A flat list of chats gets hard to scan once it passes a dozen or so.
2. Send work where you point it. You choose the destination and the request goes straight there. Automatic routing by a model can come later, as long as it is never the only way.
3. Leave each agent's session alone. The conversation should stay in the tool's own session store, where the tool can resume it. If the workspace holds the only copy, you are locked in.
4. Stop agents from colliding, with one writer per folder or one worktree per task, enforced by the app.
5. Make review the main screen. Show the result, the checks and what changed, with accept and request changes next to them.
6. Keep running when the window closes. A background process should own the agents, so a UI crash or restart doesn't stop the work.
7. Use the plans people already pay for. Run the official CLIs on the user's own login, never copy credentials, and never resend history the tool already caches.

What a workspace should not do

• Rebuild the agent. The vendors improve their agents every week, and a wrapper that reimplements tool calls or prompts falls behind fast.
• Replay chat history. Summarizing a conversation and sending it back each turn throws away the vendor's prompt caching and costs tokens.
• Hide what it adds. Any text the workspace puts in an agent's context should be counted and shown.
• Demand a hierarchy on day one. Sub-projects should be there when the work grows, and optional until then.

Workspace, editor or the agent's own app?

Questions to ask before you pick one

• Does it use my existing logins, or does it want an API key and bill me for tokens?
• Where do my conversations live if I stop using it?
• What stops two agents from editing the same folder?
• Does it keep running when I close the window?
• Does it add text to every turn, and does it show me how many tokens?
• Can I review a result without reading the whole transcript?

How Zoltbot answers those

We build Zoltbot, so read this part with that in mind.

• Claude Code, Codex, Cursor and Antigravity run through their own CLIs on your existing login. Zoltbot never asks for your password or copies your credentials, and secrets stay in your OS keychain.
• Each tool's own session store holds the conversation. Zoltbot keeps an index, never the only copy.
• A task can run in its own git worktree beside other tasks. Tasks that share a code folder run one at a time, and the rest wait with a visible reason.
• An always-on background process runs the agents, so closing or restarting the app doesn't stop them.
• Zoltbot never resends chat history. New chats get the project's instructions as bounded context, and every added token is shown.
• A finished task shows its result. Accept it, or request changes and the same task runs another revision in its own conversation.

Still planned: routing rules, an optional chief of staff that routes requests for you, pipelines and scheduled routines, and support for Gemini CLI, Grok, OpenCode, Pi and any agent that speaks ACP.