Development ·
Self-HPO jointly optimizes coding agents and their harnesses
A published framework alternates failure-driven harness search with reinforcement learning of the coding model.
Why it matters
Training the model and its surrounding tools together could change how teams improve agents after collecting real failure cases.
What to watch
Watch for reproductions, training-cost breakdowns and tests on different models and task distributions.
Still uncertain. The reported gains come from the authors’ experiments. Independent reproduction and the cost of adapting the method remain unestablished.
Impact assessment reviewed 2026-10-20.
- Who it helps
- Researchers and teams building coding-agent training systems.
- Availability
- Paper, experiment code and released artifacts are public. The repository describes a research training setup with remote execution infrastructure, rather than a ready-made desktop assistant.
A loop driven by failures
The current model proposes harness changes from failed tasks. After selecting a harness, reinforcement learning updates the model; difficult tasks feed the next search round.
Self-Evolving Coding Agents via Harness-Policy Optimization ↗Self-HPO author explanation ↗
What improved
The authors report an average resolve rate of 64.8%, up from 57.8%, across three coding benchmarks. The result is specific to their model, training and evaluation setup.
Self-Evolving Coding Agents via Harness-Policy Optimization ↗
What you can do
- Compare the published harness changes and ablations before considering a training experiment. Open reference ↗
Where the evidence stops
Training and benchmark runs were not reproduced. An author’s X thread supplies launch context, with one creator counted across its replies.
Sources read. Sources checked October 20, 2026 UTC. No external product workflow, model training or inference was executed.
Sources & further reading
Paper · paper
Self-Evolving Coding Agents via Harness-Policy Optimization ↗
Original research statement for the method, reported results and evaluation scope discussed in this article.
Wei and colleagues · page checked 2026-10-20 · Published 2026-10-16
GitHub · implementation
Self-HPO research artifacts ↗
Project-owned implementation or interface showing availability, requirements and practical constraints.
Microsoft Research Orchard · page checked 2026-10-20
X · official
Self-HPO author launch thread ↗
Original author announcement with publicly visible engagement. Same-author replies are counted as one contributor.
Zhepei Wei · page checked 2026-10-20 · Published 2026-10-19
6,342 views · 118 likes · 12 comments · 34 reposts · observed 2026-10-20 08:16 UTC
Web · analysis
Self-HPO author explanation ↗
Creator explanation of the mechanism and its limitations, alongside the original paper or implementation.
Zhepei Wei · page checked 2026-10-20 · Published 2026-10-19
Coverage note: The source bundle is a bounded sample. Independent replications and matching original discussion across all platforms were not established.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
2026-10-16: A published framework alternates failure-driven harness search with reinforcement learning of the coding model.
Assessment: Training the model and its surrounding tools together could change how teams improve agents after collecting real failure cases.