Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-04 · high
Long-horizon agent skill looks like a transferable habit, not domain knowledge, and that changes what data is worth collecting to train it.
Why now: The paper posted to arXiv in August 2026 under ID 2608.01604.
Researchers post-trained Qwen3.5-122B-A10B on 363 office tasks containing no code at all, then watched its score on SWE-Bench Pro climb 5.8 points, per the arXiv paper.
That's a jump on a benchmark built to measure whether an agent can finish real software engineering tickets. None of the training data involved writing code. For anyone building or buying coding agents, that points at a different lever than syntax knowledge. It's the ability to carry a task through many steps without losing the thread.
The authors call this goal-directed execution. They break it into four repeated behaviors. First, picking a goal and building the state needed to act on it. Then staying faithful to that goal across many steps, and checking the result against the real environment instead of assuming success. Training on office tasks improved the model on all four measures. The gains showed up both in the office setting it trained on and in the coding benchmark it never saw.
The paper doesn't say whether the effect holds on benchmarks besides SWE-Bench Pro, or on models smaller than 122 billion parameters. One result, one system, for now.
If it holds outside that one model and one benchmark, the scarce resource for training coding agents stops being code. It becomes any long-horizon task data you can label well, spreadsheets and email chains included. That's an argument for collecting messy office workflows instead of more GitHub repos.
The paper posted to arXiv in August 2026 under ID 2608.01604.
Each link below shares sources, entities, or timing with this story.
Shared entities / Shared topic / Earlier coverage / Tension
Both cover Bench Pro, Qwen3, SWE; overlapping topics (agentic, coding); earlier Bench Pro coverage from 2026-04-21.
Both cover Bench Pro, SWE; overlapping topics (agentic, code, coding); earlier Bench Pro coverage from 2026-06-10.
Shared entities / Same source domain / Shared topic
Both cover Bench Pro, SWE; reported by the same outlet (arxiv.org); overlapping topics (code, coding, completion).
Shared entities / Earlier coverage / Tension
Both cover A10B, Qwen3, SWE; earlier A10B coverage from 2026-03-16; pushes against this story (vs).
Shared entities / Shared topic / Earlier coverage / Tension
Both cover Bench Pro, SWE; overlapping topics (against, coding); earlier Bench Pro coverage from 2026-06-23.
Both cover Bench Pro, SWE; overlapping topics (code, coding); earlier Bench Pro coverage from 2026-06-07.
Both cover Bench Pro, SWE; overlapping topics (agentic, coding); earlier Bench Pro coverage from 2026-05-05.
Both cover Qwen3, SWE; overlapping topics (code, coding); earlier Qwen3 coverage from 2026-04-23.