One prompt turns Opus 5.5 into a task graph so agents only need approval on the last step

CurieuxExplorer · x · 2026-11-06

After Andrej Karpathy released autoresearch—an agent that runs experiments in a self-directed loop—@HoodieFade packaged the method into a single prompt that turns Opus 5.5 into a graph-based task system.

How it works:

Writes a check that can actually fail, first
Splits the task into parts and runs independent parts in parallel
Loops each part until it passes its check, returning only failures
Converts every mistake into a rule for the next iteration
Calls the human only for irreversible steps

The companion article, "Loops & Graphs," cites Bun's port from Zig to Rust (750,000 lines of Rust, 11 days from first commit to merge) to argue the real bottleneck isn't agent endurance but human review overhead: agents can already work for hours unattended, yet people approve every step because they have a loop but no graph. The full prompt is in the comments, with tuning guidance in the article.

Original post →

One prompt turns Opus 5.5 into a task graph so agents only need approval on the last step

More from coding & agent