My Agent Work Is Scattered Across Terminals and Chats

Codex in one terminal, Claude Code in another, a plan in a chat tab, decisions scattered through Slack threads and scrollback. Each individual session works — the agents are genuinely good — and the whole leaves nothing behind: no shared record of what was decided, what shipped, or what needs you next.

Scaling that setup means more tabs. It does not mean more capacity, because the bottleneck was never the agents. It was that the work has no home.

Work needs a durable object, not a session

The unit that fixes this is the task: a durable record that outlives the session that produced it. In Task Machine every piece of work is a task with a description, an owner, a status, and a timeline of everything that happened — human comments, agent runs, questions, approvals, and the final result in one ordered history. The chat where work starts, the schedule that triggers it, the agent that executes it — all of it creates or attaches to a task, so there is always one place to look.

Your agents do not change. Codex, Claude Code, Opencode, and the rest keep running on machines you connect — the same tools, the same subscriptions — but their runs write into the shared record instead of vanishing when the terminal closes.

One inbox instead of a tab patrol

The second half is attention. Instead of checking each terminal and thread to see what needs you, everything that requires judgment — an approval, a question, a finished draft, a failed check — lands in one inbox. You act on it in a click, and the decision goes onto the task's timeline where the next person or agent can see it.

That is the three-surface workflow: chat to direct, inbox to approve, tasks to steer. Three surfaces, one set of work, zero archaeology.

The compounding part

A system of record is not just tidier — it compounds. Agents pick up context from the task history instead of starting cold. Corrections persist as memory. And when you ask "what actually happened with the launch email?", the answer is a timeline, not a reconstruction from four scrollbacks and a guess.