Sierra's new open benchmark asks the next question
sierra's new open benchmark asks the next question: can an agent build an agent? on hyper-tau-bench the best solo setup, claude opus 5 in claude code, passes 23.9% of tasks while codex with gpt-5.6-sol hits 22%, and the same models paired with an engineer reach 82.2%.
the gap is still the human in the loop.
Context
Sierra's blog of 8 September 2026 describes hyper-tau-bench, an open benchmark where a developer agent in a sandbox receives a simulated business's records and a simulated client, builds a customer-service agent, and is scored on held-out simulated-user tasks. Its leaderboard lists Claude Opus 5 in Claude Code at max reasoning on 31 August 2026 with 23.9 percent, GPT-5.6-sol in Codex at xhigh with 22.0 percent, and a Human performance row at 82.2 percent. Sierra writes that the same class of model, paired with an engineer with deep context, reaches 82.2 percent. One finding on the page is that builds which asked the client more questions scored higher.
The benchmark and leaderboard are Sierra's own, not independently run. The 82.2 percent is a human plus model reference, an engineer-built agent, and Sierra says the same class of model; the note's the same models paired with an engineer is not a controlled pairing of those exact models. Scores are pass rates on held-out tasks in this benchmark only. The arXiv page was read as a snippet only. That the gap is still the human in the loop is the author's reading.
Watch next
- Leaderboard changes after 8 September 2026, the paper's task count, and independent replications.
Sources
- Evaluating agents that build agents (Sierra, 8 Sep 2026)sierra.ai
- hyper-tau-bench leaderboard (Sierra Research)sierra-research.github.io
- hyper-tau-bench paper (arXiv 2609.04611)arxiv.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 03:19 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →