ChatGPT Work and Codex alternative for teams that want a model-independent operating system for agent work

ChatGPT Work marks a clear move toward task-first agent work: the new Task action emphasizes giving an agent an outcome and letting it work for longer than a normal chat response. LatchLoop has been built and used around tasks since February 2025, but the task model is different. A Work or Codex task remains primarily a long-running agent thread started by one person; a LatchLoop task is an editable living document inside a shared project and can change across multiple agent runs and teammate contributions.

ChatGPT Work turns outcomes into long-running tasks across apps, files, plugins, the browser, and—on desktop—computer use. Codex is the software half of the family, with parallel local or cloud threads, managed worktrees, CLI and IDE surfaces, GitHub and remote SSH environments, diffs, tests, mobile Remote, scheduling, and review tooling.

LatchLoop is an all-in-one, multiplayer workspace for coding and general agents: an agent-native editable task is the shared source of intent, while the built-in editor and terminal, preview and element inspector, diff and pull-request review, PR questions and change requests, direct merge controls, teammate approvals, plugins, artifacts, agent apps, and automation keep the complete lifecycle in one platform. ChatGPT Work + Codex is a direct competing app, not an app embedded in LatchLoop. LatchLoop can invoke the Codex coding harness through Agent Client Protocol, but its distinctive value is the shared project where teammates can see tasks, assign ownership, collaborate with agents together, and choose their preferred model or provider.

LatchLoop is like multiplayer for working with agents: everyone with access to a shared project can see its tasks, assign work to teammates, co-edit the brief, send attributed direction, and understand what people and agents did. Collaborative task documents stay beside agent activity, while the desktop app provides a built-in editor/IDE and terminal, preview and element inspector, diff and pull-request review, PR questions, change requests, and direct merge controls. In the same platform, general agents use plugins and skills to create artifacts and agent apps or run recurring automation.

LatchLoop begins with a collaborative task document rather than a disposable prompt. Teammates can co-edit the brief, assign an owner, use Ask and Implement Plan, attach files and links, and then choose LatchLoop’s harness, Codex, or Claude Code. Attributed messages, visible agent activity, editable to-dos, and the persistent task create a durable paper trail of what people asked for, what the agent did, and why the result changed.

For web and mobile coding tasks, LatchLoop runs cloud agents deterministically confined to the task’s assigned branch. That reduces overlap and unintended cross-branch edits, at the cost of less freedom than a broadly authorized local agent. Local agents can receive approved broader permissions, and the document editor can push to main. Until native local worktrees are available, LatchLoop recommends one local agent per project and parallel cloud runs for additional tasks.

LatchLoop is designed for portability. Teams can export the full prepared prompt, choose supported model providers without token markup, switch between the LatchLoop harness, Codex, and Claude Code, and keep general-agent memory, knowledge, processes, and SOPs as files in a customer-owned GitHub repository. Those process files remain inspectable and reusable with another harness.

Pilot the families with one connected knowledge task, one ambiguous software change, one scheduled workflow, and one task steered from mobile. Use the same source material and approval policy.

Score completion quality, but also score task preparation, parallel isolation, team attribution, review, process portability, realistic monthly cost, and whether a colleague can reconstruct what happened.