StateM Framework Reaches 95.3% Accuracy on Terminal-Bench via Harness Scaling Without Model Weight Changes
liuziwei7 · x · 2026-08-18
The paper introduces StateM, an agent-native runtime that improves agent execution accuracy via Harness Scaling (scaling the control flow around the model) without modifying model weights. It organizes execution around durable states, phase-local context, checked transitions, and recoverable runbooks.
- Core Conclusion: Combining human direction with agent action yields better workflow control than local optimization or pure human effort.
- Performance: On Terminal-Bench 2.1, StateM boosts GPT-5.5 xhigh to 92.1% (vs. 83.1% baseline). With GPT-5.6 Sol xhigh, it achieves 95.3% raw accuracy at a cost of $15 (vs. $574.68 for the GPT reference).
- Transferability: Runbooks transfer seamlessly across models; e.g., it raises DeepSeek-V4 Flash to 88.8%.
- Availability: Enables open-source models to match frontier proprietary models performance without training and with zero transfer costs.
More from coding & agent
- Permix: Lightweight Type-Safe TS Permissions Library Wins Developer Praise — jonathan_wilke · 2026-08-18
- Give agents a centralized source of truth before they hallucinate — eptwts · 2026-08-18
- After letting her agents queue up commits, her GitHub daily streak exploded to 2026 — christine_hall · 2026-08-18
- Why logging is vital for agent apps: accountability, evals, and self-healing — jasonkneen · 2026-08-18
- Supabase ships battle-tested UI components — hand them to your agent — dshukertjr · 2026-08-18
- PyLate hands model maintenance to SentenceTransformers' Tom Aarsen — antoine_chaffin · 2026-08-18