--- title: 'In the News: September 13, 2026' description: 'Ryan Lopopolo argues that no grader is unhackable, exposing the risk of trusting agents in domains where users cannot judge the work.' canonical_url: 'https://darkfactory.dev/news/2026-09-13-morning' markdown_url: 'https://darkfactory.dev/news/2026-09-13-morning.md' collection: news date_published: '2026-09-13T07:15:00-04:00' date_modified: '2026-09-13T07:15:00-04:00' --- # In the News: September 13, 2026 Ryan Lopopolo, a harness engineer who has spent years building the layer that keeps coding agents on track, argues that grader hacking cannot be patched away. Every measurement system reflects someone's judgment about acceptable shortcuts. The risk becomes hard to assess when people deploy agents in fields where they cannot spot mistakes themselves. A second item connects familiar reproducibility practices to long-running agent context. Tests, commit history, repository structure, project instructions, and decision records can give an agent the continuity it needs to stay coherent over months of work. ## 1. There is no such thing as an unhackable grader, a veteran harness engineer argues **[Aligned to whom?](https://hyperbo.la/w/aligned-to-whom/)** · Ryan Lopopolo, harness engineer, formerly OpenAI, now at Google Cloud · hyperbo.la, September 12, 2026 Lopopolo starts with a practical limit: experts can catch an agent's mistakes in their own fields. Outside that expertise, in his examples double-entry accounting, finance, law, and operations, users rely entirely on the model's priors without knowing whether those priors are sound. He is blunt about his own field: "I am an expert software engineer and I am not happy (and never have been) with the default behaviors of the model when producing software." His larger claim concerns grading itself. Lopopolo argues that every rubric, eval, and human rater that shaped a model's behavior brought its own blind spots, which the model's priors then inherited. "There is no such thing as an unhackable grader," he writes, because people disagree about what counts as a permissible shortcut. He connects this to a problem this feed has tracked before: models trained without a memory of past mistakes have no equivalent of "a fear of future regret." That helps explain why long-term coherence in agent-produced systems remains, in his words, "a very unsolved problem." **Why it matters:** Before handing an agent unsupervised work outside your own expertise, ask who wrote the rubric it is optimizing against and whether that person's blind spots match yours. ## 2. A researcher's case: your reproducibility habits are already agent context **[Reproducibility in the Age of Agentic AI: Context Engineering at the Timescale of a Codebase](https://arxiv.org/abs/2609.11728)** · Lorena A. Barba, George Washington University · arXiv, published September 2026 Barba, a longtime reproducible-research advocate writing on agentic coding for the first time, maps five familiar software artifacts to the context a coding agent needs: a test suite, a clean commit history, a well-organized repository, a README and conventions file, and a decision record. In her framing, tests become executable semantic memory and commit history becomes a form of episodic memory. Her central line: "Reproducible research practices are context engineering for AI coding agents." She supports this with what she calls a convergence argument. An unrelated large-scale study of AGENTS.md files and a research-software-engineering community document, written for different reasons by different people, arrived at nearly the same list of artifacts. The paper also examines skill erosion and the risk that maintaining artifacts for an agent's benefit could quietly change what a team optimizes for. **Why it matters:** A decisions log or conventions file can directly improve an agent's consistency. Teams should account for that value when deciding whether the upkeep is worth the effort. --- ## Also this cycle - **[Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents](https://arxiv.org/abs/2609.11677)** · arXiv (Yue, Cui, Sun and 11 coauthors), published September 10, 2026 · A training method for self-evolving coding-agent harnesses that separates model-specific quirks from real harness bugs before rewriting anything, reporting a 1.84x training speedup and an 18.56 percent accuracy gain over prior methods on airline and retail agent benchmarks. This corpus already tracks several papers in this specific sub-field. The numbers refine an existing approach. --- ## Thread watch _Discussions gathering force. No primary read yet, so these are reported as discussions, not as findings._ - **[Garry Tan on X](https://x.com/garrytan/status/2098666551629267324)** · X · Tan's quote-tweet was reported as saying that founders outside hardware are converging on building a domain-specific harness instead of remaining a system of record. The post had reached roughly 685,000 views, 3,400 likes, 269 reposts and 137 replies at 19 hours after posting. By the next morning, a separate trending topic tied to the same line had drawn more than 2,300 posts. The post itself and the underlying Demo Day observation have not been independently verified.