An experiment that exploits a weakness of AI coding agents — tests pass, but nothing gets fixed — new routing research that moves skill selection inside the model, and a PR with direct implications for power efficiency in local LLM operations. From today's news, here are six stories with practical payoff for developers, plus one quick aside.
All the Tests Passed — But the Bug Wasn't Fixed
An experiment published on GitHub — and just starting to pick up traction on Hacker News — found that when 26 AI agents were set loose on code with a single planted bug, every one of them passed the tests, and every one of them left the bug in place.
At collection time we haven't been able to verify the repository's detailed experimental protocol, but the core claim is simple: the agents could produce changes that pass the existing test suite, yet never touched the root cause of the planted bug.
That throws down a question for the agent era: how should the assumption that "green tests means fixed" change? For teams delegating fixes to agents on legacy codebases with thin test coverage, this experiment looks like a good catalyst for rethinking how results get evaluated.
Gavel: Skill Selection Inside the Model
Load an agent up with too many tools and deciding which skill to use, and when, gets hard. The prevailing answer is progressive disclosure — loading skill metadata into context and letting the model pick — but as skills multiply, they eat more and more context.
Gavel, presented via Hugging Face Daily Papers, takes a different route: it selects skills using nothing but a frozen LLM's own hidden states and judgments. The final skill is decided in two stages — "Glance," which narrows the candidate pool via hidden states, and "Verdict," which combines the same LLM's likelihoods with yes/no judgments. The backbone is never trained; the additional trainable parameters total just 7.9 million on Qwen3-32B, and registering a new skill takes a single frozen forward pass.
The paper also introduces SkillTraj, a new benchmark that measures routing ability amid the noise of agent trajectories — research squarely aimed at the practical pain of running agents whose skill libraries keep growing.
Combine Continual Learning Mechanisms and Memory Lasts 28× Longer
Also drawing attention on Daily Papers is work on catastrophic forgetting. A simple recipe that combines multiple continual learning mechanisms dramatically improves long-term retention: average final retention across three datasets reached 34.9%, up from 1.2% — 28× that of sequential SFT.
The result — getting this far not by going all-in on one technique but simply by "compositing" existing mechanisms — offers a useful reference for memory design in agents built for long-term operation.
A LACT PR That Breaks the RTX 5090's 400W Floor
A PR has landed on LACT, the GPU management tool well known to the local LLM crowd, that pushes past the minimum power floor set by the VBIOS.
With the stock VBIOS, the RTX 5090 is pinned to a 400W minimum, but reports say that with this change, dropping to around 300W costs only a sliver of text generation performance. On the RTX 6000 PRO MaxQ, it apparently ran at as little as 75W. Controllability of shunt-modded cards should improve as well.
On an always-on local LLM server, power draw goes straight onto the electricity bill. This PR, which makes clock and voltage control more flexible, is a practical change that translates directly into lower costs for 24/7 operation.
An MCP Experiment Wiring a Local LLM into GIMP
A setup reported on LocalLLaMA combines llama.cpp's MCP server functionality with "gimp-mcp" to hook the local model Qwen3.6-35B-A3B into GIMP. The candid verdict on the first prompt — "draw a picture of a flower" — was "still rough," but the setup itself is reportedly painless: one JSON config file plus an extra launch option.
That an image editor can actually be driven from a local LLM is a sign MCP's reach is spreading beyond code execution.
Weathernews Tackles "External AI Is a Leakage Risk" by Building Its Own
On the Japanese side, an article reports that Weathernews — wanting AI that's easy to weave into daily operations while keeping data-leak risk in check — took the in-house development route, and adoption has spread to the point where more than 80% of its employees use it.
The same day, another article appeared in which practitioners at HP Japan discuss their decision framework for where to run in-house AI workloads: cloud or local. Picking the execution environment based not only on model performance but also on data confidentiality and operating cost — the discussion at Japanese companies is clearly getting concrete.
Bonus: What Makes LLMs Laugh
According to a laugh.so research report, agreement rates when LLMs judge each other's humor vary widely — anywhere from 54.9% to 92% — but nearly all models favor absurdist material and roast-style jokes. Gemini was the most consistent judge, though still less consistent than humans. Before you hand judging duties to an LLM, you might at least be able to agree on what counts as funny.