AI NEWS TODAY

The BriefThree separate pieces landed in the last few days that all point at the same thing, and it is not the models. Steve Yegge published part one of The Continuous Thunderdome describing how Gas Town, his reusable harness for running coding agents, in his words "fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly." The new version picked up a "just two more things" tic and would never declare the harness done, so it kept fiddling with the harness instead of doing the work. He did not fix it. He concluded a reusable framework is the wrong shape and that the harness has to be welded to the specific application. Lilian Weng, separately, defines the harness as everything around the model that decides how it plans, calls tools, manages context and checks its own results, and argues the real improvement now comes from that layer rather than the weights. Sean Goedecke adds the uncomfortable half: the human is the bottleneck, because the hard part is saying precisely what you want. If you run agents on real work, the practical takeaway is that a minor version bump can change your agent's behavior enough to break a workflow that was fine for months, and it will not look like a failure. It looks like work that never quite finishes.

Level UpPick one place you forward AI output to another person without reading it first, then stop doing that this week. Niklas Gruhn calls the habit being a meat proxy: pasting a model's answer onward and letting the next human do the thinking. His fix is four steps and the last one is the test. Read it, understand it, check it is actually true, then write the response in your own words, because writing it yourself is decent proof you did the first three. The cheapest version for a working week: the next time you are about to relay a model's summary of a ticket, a call or a contract, type the one sentence that matters in your own words instead of pasting the paragraph. Don't be a meat proxy