uygnoey Delegating whole tasks to agents that run unattended
Yusuke Kaji, General Manager of AI for Business at Rakuten, has tested Claude models since September 2024 — across nearly a dozen model launches — and describes Claude Fable 5 as a step change for long-running enterprise agents. The change he points to is not raw speed. It is that the model checks its own work as it goes, which is what makes an unattended multi-hour run worth starting.
This skill describes the delegation pattern as the post presents it.
Instructions
1. Understand why long runs used to fail
Before Fable 5, setting an agent loose on a multi-hour task without human oversight was a gamble. Kaji describes the failure directly: "If they choose the right path in the first step, everything is fine. But if they choose the wrong direction in the first pass, the agent spends significant time to fix the path, or even fails to reach the destination."
On a job meant to run five hours or a full day, one early wrong assumption could burn the entire run, and the only way to catch it was a person checking in.
The failure mode was a lack of self-verification. Any model can take a wrong first step. The problem with earlier models was that they did not check their own work as they went, so an early wrong turn went unnoticed, compounded over the run, and produced a suboptimal result hours later.
2. Check for the three behaviors that make an unattended run viable
Kaji's team cite three behaviors that distinguish Claude Fable 5 from its predecessors. Before you extend a run's length, confirm the model actually does these things on your work:
- It re-checks its own assumptions. When the state of the task changes midway, it notices and corrects a wrong assumption before acting on it, rather than committing to a bad path and discovering it hours later.
- It returns to first principles at each step. It re-validates against the original intent without being told — the course-correction Kaji used to have to make himself.
- It matches the team's taste. Even with minimal guidance, its judgment on ambiguous calls lines up with theirs. Kaji coined a term for this: taste alignment.
The full description of each behavior, and what the post does and does not specify about it, is in references/self-verification-behaviors.md.
3. Raise the unit of work from the chunk to the whole task
"Before Fable, we had to break work into well-defined chunks for the agent to execute," Kaji says. Now he can hand over a whole task and run several at once.
This is the practical shift. Stop decomposing the job for the agent as a defensive measure against early wrong turns, and hand it the outcome instead. The corollary is that you can run several such tasks in parallel, because none of them needs you watching it.
Brief the task with templates/stretch-task-brief.md.
4. Prepare stretch tasks deliberately when a new model arrives
Kaji likens testing a new model to embarking on a "new quest," and describes a deliberate practice: "The way a good leader prepares stretch goals for their people, we prepare stretch tasks for a new Claude. Maybe Claude is nudging us to stretch, too."
Keep a standing set of tasks that the current model cannot reliably finish, and point each new model at them. That is how you find out whether the unit of work you can delegate has grown.
5. Give agents memory across runs
"Our agents with memory remember what went wrong in past sessions and avoid repeating those mistakes."
Memory between runs is part of what makes long autonomy compound rather than repeat. A run that fails in a particular way should make the next run less likely to fail that way.
6. Expect human judgment to become the constraint
Rakuten's agents close issues roughly 10x faster across every domain, and the number of tasks the organization takes on keeps rising. Adding more agents does not add judgment. The faster the agents run, the more the organization's progress depends on a person closing the loop.
Because the model self-corrects mid-run, sign-off becomes feasible for the first time, and the unit of work delegated shifts from the task to the decision. The absolute number of tasks keeps climbing, but the ones that truly need a human stay at a focusable level.
Kaji names not having to jump in and steer mid-run as the biggest productivity win of all: it lets his team spend its time on the decisions only people should make, and keeps an AI-native organization accelerating instead of stalling on human course-correction.
Examples
The Rakuten runs as the post reports them — the overnight job, the parallel whole-task delegation, and the 2 a.m. self-correction — are in examples/rakuten-agent-runs.md.
The line Kaji uses to describe the difference:
"We tested Fable, and we love its capability for self-reflection and self-verification. Compared with previous models, it understands its mistake before I point it out at 2 a.m. or 3 a.m. — so that I can sleep."
Source
Working at the frontier: How Rakuten builds agents overnight with Claude Fable 5 — Claude blog, July 20, 2026.