Agent actions keep getting cheaper to produce, and the most useful builder posts of the day control the action itself, at decode time or at the sandbox or the push, instead of trusting what the model says it will do.

Agent actions keep getting cheaper to produce, and the most useful builder posts of the day control the action itself, at decode time or at the sandbox or the push, instead of trusting what the model says it will do.

Model calls keep getting cheaper, and three of the weekend's most-discussed agent projects spend their effort on what the model sees instead.

AI now turns up bugs and claims faster than anyone checks them or ships the fix, so the useful work this weekend sits in the second step.

This weekend's AI stories grade a token by what it buys, and the harness, the hardware and the balance sheet each give a different answer.

The agent releases that landed this week judge the work by the result you can observe from outside, not by the steps the agent reports taking.

Two frontier labs disclosed their own models overreaching on live networks the same day, and the containment response in both converged on isolating the network.

Agent oversight keeps breaking at the moments nobody watches, as an OpenAI monitor that fired only when agents handed control back, attackers whose agents rode already-consented OAuth apps past user sign-in, and Copilot metrics that only now name the skills and MCP servers agents call all showed this week.

Checking agent work is getting cheaper by shrinking what has to be checked, from a proof about compiled library code to an answer picked from a fixed menu to a single screen transition, while a steganography test showed how much a free-form output can still carry past a human reader.

Most broken tool calls from a self-hosted agent model are a serving setting rather than a model flaw, because vLLM only forces tool calls…

Impeccable's most useful part is not its design advice but 62 deterministic detector rules that catch generic AI frontend patterns with no…

A design spec is only done when a reviewer that never saw the conversation can follow it, and EGM turns that test into a command you run…

Long agent sessions now fail by holding on to stale plans, so the cheapest fix is a planned reset that keeps the work on disk and throws…

GVS5H's coding gains come from bounded fresh model calls that keep their state in small files on disk, a pattern any agent setup can copy,…

DeepOpen can replace fixed-choice LLM calls with a 421M-parameter encoder that answers in milliseconds, but its own reports say the…

Broccoli2Shell lives in a decoder bundled inside the runtime, so the fix is a Node or Python wheel version rather than a package bump, and…

AI text is harder to read because post-training rewards correct answers over clear ones, so builders should ask for and check readability…