Kyle Baxter

At least for this project, I’ve been indefinitely productive with using 4.5. This last spring, with 3.7 and Gemini Pro 2.5, that period of productivity would hit a wall at some point of codebase complexity. It’s all anecdotal, but it feels quite different.

Sonnet 4.5 tbh has been extremely good for a recent project for me. Not only is it acting very agentically—taking a significant number of steps in response to a request from me, with each step being quite good—but the apparent depth of thought behind the actions it takes feels better.

My POC codebases used to reach a point where Cursor’s agent mode would make a mess of them; I’m not hitting that point as easily now.

Your note about model improvements making this realistic now is good. I wasn’t able to do more AI-first dev experimentation since the spring until these last couple weeks, and GPT-5 and Sonnet 4.5 are remarkably better.

Miller’s apparent elevation to shadow president and his seeming intent to declare an end to the constitutional order would, in any other administration, be treated as a executive branch meltdown that was immensely embarrassing for the president and requiring Miller’s immediate removal

Letta Cloud now provides Anthropic's `memory` tool, enabling agents to dynamically manage their own memory architecture. Agents can create, delete, view, and modify memory blocks as they learn.

In the video below, we use the memory tool in the ADE with Sonnet 4.5 to create a new memory block.

I am so here for Chevelle Bluesky. My dad had a ‘70 SS 396 when I was a kid. I miss that car. His first car was a Chevelle, too, and I think it was a ‘67.

yes, yes you did. beautiful car

did you just casually drop a beautiful ‘66 or so Chevelle

one way you could describe lincoln's approach is of never getting ahead of the public on its attitudes, but holding an immovable set of principles on where the country ought to be and where he hopes to lead it

The protocol is the API.

I highly recommend this post. Bluesky is great. atproto is the real star.

bit by bit, the administration is laying the groundwork and actually carrying out a system of repressing democratic life in america. redefining opposition as terrorism, targeting critics, investigating civic funders.

we're so far beyond normal and, again, only eight months in.

The Biden administration didn’t force social media companies to do anything, even if what they did do—submit accounts and posts they thought violated the social media company’s TOS—was itself noxious. Comparing the two is a category error.

David Letterman Laments Jimmy Kimmel Suspension: ‘This Is Misery’ Legendary late-night TV host David Letterman slammed ABC's decision to suspend Jimmy Kimmel.

“'You can't go around firing somebody because you're fearful or trying to suck up to an authoritarian criminal administration,' Letterman says” variety.com/2025/tv/news...

There is no “critics say”; the President is demanding his DOJ go after his enemies.

Nine Inch Nails fucking kicking ass

Gonna use social media for its real purpose for a second (post blurry ambiguous show photos)

revealed preference is that many US elites prefer crony capitalism and promises of a mechanical god to science and innovation

That matters for a lot of reasons, but one is it makes the whole debate about what is and isn’t an agent irrelevant. Who cares. If something does what it says it does and that’s what you (being a user or another agent) needs, how is irrelevant. It’s the protocol that matters, not the how.

That’s a nice advancement, because it allows us to loosely couple components and compose them. How each component works is irrelevant; what matters is it does what it says it does and conforms to a common messaging protocol.

The part of a2a I find especially meaningful is the agent card, because it extends the agent abstraction. If an agent specifies what it does and how to use it, then what’s in the agent black box (how it works) is an implementation detail. The agent card and interface is what matters. How doesn’t.

But that is where the terminology becomes a little problematic. Similarly, an agent can be a single unit (model+instructions+tools), or it could be multiple agents with control flow—and you’d refer to the latter as an “agent,” too. Not sure if we need a new term for single unit agent or for multi.

Tool calling is also helpful as a concept, though I think the abstractions break down a bit here. A tool call can invoke a tool; but it could also invoke another “agent.” Is that a tool call? Is it agent communication? It doesn’t really matter; it’s a model calling a function with a purpose.

Similarly, a tool represents an atomic action which can be taken. It is lower level than an agent; a single tool could be useful for a wide variety of agents. But it clarifies how to break problems down and how agents can compose tools to accomplish a goal.

“Agents” and “tools” is more a conceptual advancement than technical. Agent helps because it formalizes thinking about actors with particular roles they perform, which is a clarifying conceptual abstraction for how to build AI capabilities. Each agent has a role it performs which can be composed.

Unfortunate indeed.

Is this satire? I genuinely can’t tell anymore because what’s real has become so absurd.

The sirens can’t be any redder or blare any louder

This was after Colbert was canceled. Which was two months ago.

iOS and iPadOS 26: The MacStories Review Old and new through the liquid glass.

www.macstories.net/stories/ios-...