How we run dozens of coding agent sessions on one Taskhold board
By Jake Marsh, who builds Taskhold. Updated 2026-11-04. Facts about other tools checked 2026-11-04. 3 min read.
Short answer: Great Work is one person and a lot of coding agent sessions. Every session files its work on one Taskhold board: in our first eight days that was 715 tasks, split into 2,331 steps, across 49 projects, with 55 running at once at one point. One rule in the global CLAUDE.md makes every agent do it, and the live board replaces scrolling through terminals.
We build Taskhold, and we run our own studio on it. Great Work is one person plus coding agents working on dozens of products at the same time: web apps, Mac apps, add-ons and the sites that sell them. Since 2026-10-28 every Claude Code session on that machine reports into one Taskhold account. The numbers below are from that account, read from our own database on 2026-11-04.
This is our setup, not a benchmark. If you run one agent in one repo, Claude Code's built-in task list is enough; see Claude Code tasks vs Taskhold.
The numbers after eight days
| Item | Count |
|---|---|
| Tasks filed (all by agents) | 715 |
| Steps those tasks were split into | 2,331 |
| Tasks finished | 442 |
| Projects (one per repo folder) | 49 |
| Distinct agent labels (client plus repo) | 44 |
| Tasks in progress at the time we counted | 55 |
| Times an agent stopped to ask a question | 28 |
Our own account, 2026-10-28 to 2026-11-04. No customer data.
Every task was filed by an agent, none by hand: when Jake asks for something in a session, the agent puts it on the board before it starts.
The one rule every session follows
A short section in the global ~/.claude/CLAUDE.md applies to every session in every repo (the pattern is in the CLAUDE.md task workflow). The agent files each request with steps, completes steps as it goes, blocks with one question when it needs a decision, and finishes with a summary of what changed.
# tasks (Taskhold)
Use Taskhold for all work. I watch the live board instead of scrolling chat.
- always pass project: "<repo folder name>"
- every request goes on the list: taskhold_begin with title, project, steps and
agent: "Claude Code in <repo>"
- each step done: taskhold_complete with start_next: true
- long step: taskhold_progress every few minutes, with a plain-sentence message
- waiting on me: taskhold_block with one clear question
- done: taskhold_complete on the task with what changed and how you checked itProjects are matched by name, so an agent working in a new repo creates its project on first use. Nobody sets anything up per repo.
What the live board replaced
- Switching between terminals to see which session was still working. The live board shows every running task with its progress bar, the step it is on and its last message.
- Questions lost in scrollback. A blocked task moves to the top under Needs you with the agent's question and suggested answers, and gets answered from a phone.
- Asking an agent what it did. The finished task carries a summary of what changed and how it was checked.
What still goes wrong
- Progress is self-reported. An agent that says 80% may be stuck; the board shows when each one last reported, so a quiet session stands out.
- Sessions end without closing their task. Some of the 55 in-progress tasks belonged to sessions that stopped mid-way; the morning summary lists agents that went quiet so they get closed or picked up.
- Steps vary in size. One agent plans three steps where another plans ten, so a percentage is a rough guide, not a clock.
See how far along your coding agents are.
$12/month or $120/year after the trial. Card up front. Cancel before day 7 and you pay nothing. Every agent you run is included.
Questions
- Do the agents need to be told to use Taskhold every time?
- No. The rule lives in the global CLAUDE.md, so every new session reads it before the first request. Per-repo setup is not needed because projects are matched by folder name.
- Can I see the board?
- The public demo at /demo/live uses the same board components with sample tasks. Our real board has private project names, so we publish counts only.
Related
Updated 2026-11-04. Something wrong or out of date? Email hello@greatwork.company.