- Who it helps
- Researchers training agents that use tools.
- Availability
- November 3 paper and Apache-2.0 implementation are public. The first code release focuses on AppWorld.
How the retries work
An execution setup, or harness, includes prompts, memory and tool access. A fixed pool offers alternate setups when every attempt under the original setup fails; the trained policy can return to the original setup.
Paper and release differ
The repository includes twenty AppWorld harnesses. Public recipes use validation-based selection with behavior profiling and model voting disabled by default. Rebuilding the offline harness search is outside the launch commands.
Harness-S trains from alternate agent execution setups ↗Harness-S implementation ↗
What you can do
- Read the linked primary evidence and its access limits before planning a trial. Open reference ↗
Where the evidence stops
No training or benchmark reproduction executed. Results from the paper do not establish equivalent support for every environment in the first release.
Sources read. Sources read November 6, 2026 UTC. No third-party project, model inference or complete product workflow was executed.
Sources & further reading
Paper · paper
Harness-S trains from alternate agent execution setups ↗
Primary page read November 6. Claims were checked against the page; the workflow was not executed.
Harness-S authors · page checked 2026-11-06
GitHub · implementation
Harness-S implementation ↗
Primary page read November 6. Claims were checked against the page; the workflow was not executed.
DeepExperience · page checked 2026-11-06
Coverage note: Original YouTube demonstrations and independent cross-platform attention were not established. No end-to-end workflow was executed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Harness-S retries failed training tasks under alternate execution setups and learns from those attempts.