Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-31 · high
It splits agent composition from runtime adaptation, and its GitHub repos are still active, not archived research code.
Why now: The paper posted August 28 and both GitHub repos were still receiving pushes as of August 31.
An open-source coding harness called openJiuwen scored 82.6% on SWE-bench Verified and 87.19% on Terminal-Bench 2.1, according to the openJiuwen paper posted August 28. Those numbers beat the strongest official-leaderboard scores the paper selected for comparison, by 3.4 points on the coding benchmark and 3.39 on the terminal one. Both benchmarks score whether an agent finishes a multi-step coding task, not a single prompt, so a few points there is a wider gap than it looks.
The paper splits the harness problem into two pieces. Structural Composability lets developers combine single agents, delegated sub-agents, and a coordinated swarm on one shared execution layer. Runtime Adaptivity changes while a task runs, reading semantic diagnostics, execution outcomes, and task progress to rewrite context and feedback, all under a model policy that stays fixed.
The code behind the paper isn't sitting idle. The openJiuwen-ai/jiuwenswarm repo has 6,309 stars and its agent-core companion has 420, and both were pushed on August 31.
Each link below shares sources, entities, or timing with this story.
The errors trace back to how the benchmark pairs pull requests with GitHub issues, not just to model quality.
Evolved harnesses lost to a matched-budget sampling baseline once tested on a benchmark the search never saw.
Weights and the full training stack are public, and the 35B agent beats GPT-5.5 plus Codex while capped at 12GB of VRAM.
A training-free fix called ChannelGuard held steady across three model backends, filter or no filter, blocking every tool-poisoning attempt.
ActBench ran 24,000 attack trajectories across 15 models and six harnesses; no harness pushed success below 73.7%.
A blinded judge checks root cause and impact against 95 real CVEs, and no frontier model made the ten-model lineup.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.