You’ve been using AI agents for weeks — spinning them up, watching them write code, reviewing the output. It works. But you keep losing track of what they’ve done, what’s left, and what you asked for in the first place.
The agent is productive. You’re the bottleneck. Not because you’re slow, but because there’s no system for the work between you.
The Missing Layer
AI agents are good at executing. They’re fast, thorough, and increasingly autonomous. But they don’t know what matters. They don’t know what’s next. They pick up whatever you hand them and run — which is great until you realise you’re spending more time deciding what to hand them than they spend doing it.
What’s missing is a shared surface. Something the agent can read from and write to. A backlog it can pull from the way a developer pulls tickets.
Connecting the Loop
Task Register TR-1 is a task manager built around GTD, with an MCP server that lets AI agents interact with it directly. Projects, next actions, inbox processing, weekly reviews. The full workflow, exposed as tools a language model can call.
When you connect Claude Code to TR-1, something shifts. You open a session and say “pick up the next task.” The agent queries the project, finds the highest-priority next action, marks it in-progress, does the work, commits the code, and marks it complete. In the sidebar, the dot changes — pending to in-progress to done — while the agent works.
That moment is more interesting than you’d expect.
What Changes
Not the speed. The planning.
When your agent can pull from a backlog, you start thinking differently about what goes into it. You separate the work that needs your judgment — product decisions, design direction, priority calls — from the work that doesn’t. Bug fixes. Boilerplate. Scaffolding. Content formatting. Migrations.
The backlog becomes a boundary between thinking and executing. You do the thinking. The agent does the executing. And the system keeps both sides honest about what’s been decided and what’s been done.
The Mirror
TR-1 tracks who created and completed each task — human or agent. After a few sessions, the activity view becomes a mirror you don’t expect. The agent completes tasks at a pace you can’t match. But the tasks it completes are the ones you already decided on. Your throughput is lower, but your contributions are the decisions that shape everything downstream.
That’s the right division of labour. Not human vs. machine speed, but human judgment feeding machine execution.
The Quiet Part
The most useful thing about giving an agent a backlog isn’t the automation. It’s the forcing function. You have to articulate what you want done. You have to break it down. You have to decide what’s next and what can wait.
Those are the skills that matter most when the cost of execution drops to near zero. Not prompting. Not tool selection. Planning.