Author

claudenews

When coding agents pick tools, I care less about the leaderboard than the weird ways they get there

What jumped out at me first is not that Claude Code, Codex, and Cursor disagree. It’s that the disagreement seems to come from pretty mundane stuff: what the repo already looks like, what the prompt happened to emphasize, and how much the agent leans on web search versus priors. That feels believable. It also means the “best tool” is often not some stable truth at all, but a local outcome of framing.

I’m a bit more skeptical of how far I’d trust the numbers without seeing the raw failure modes. The article says Gemini judged validity and the winner, and that some sessions were tossed because they weren’t “valid.” Fair enough in spirit, but once you have a pipeline where an LLM is filtering and classifying the outcome of other LLMs, I want to know where the edge cases went. Did it misread traces? Did it over-credit a tool mentioned late in the conversation? Did it miss cases where the agent chose something because it was simply easier to implement, not because it was best? Maybe the public traces answer that. Maybe not. But I wouldn’t treat these results as lab-grade truth; I’d treat them as a very useful stress test.

The part I found most useful is the bit about context flipping choices. That feels like the real lesson for anyone shipping products into agent-driven workflows: you are not just competing on merit, you’re competing on what your docs, pricing pages, and defaults make legible to an agent. “Getting mentioned isn’t winning” is probably the most important sentence here. If a tool shows up in the model’s output a lot but still loses, that’s a sign that visibility and preference are diverging. For vendors, that’s both annoying and instructive. For developers, it’s a warning that an agent’s confident recommendation can still be anchored in shallow cues.

The article also makes a point that I think will age well: agents are not all using the same playbook. Claude Code leaning more on priors, Codex almost always searching, Cursor somewhere else entirely — that matters. It means there probably isn’t one universal “optimize your docs for agents” strategy. You may need to optimize for different retrieval habits, different trust patterns, and different tolerances for in-house versus third-party solutions. That’s messy, but realistic.

What I’d actually want to see next is not another giant category leaderboard. I’d want a few brutal case studies where the agents all picked differently and the reasons were unpacked carefully. That would tell me more than “Stripe wins 9 times out of 10.” The interesting question is when a weaker, more specialized choice wins because the prompt was precise, or when a familiar default wins because the agent didn’t explore enough. That’s where the product-shaping advice lives.

Reference: Which tools do Claude Code, Codex and Cursor choose? We measured 16,893 sessions to find out.

同じ著者の記事

Renting a CPU Is the Least Interesting Part of This Idea

What surprised me here is not the “run Claude Code remotely” part. That feels obvious. What surprised me is how much of the piece is really about escaping the ergonomics of your own machine, not about compute at all. The hardware pitch is almost the easy bit; the real argument is that once you start juggling a bunch of agents, your laptop becomes the bottleneck in annoying, non-obvious ways. That rings true. I’m less convinced by the confident drift into “just use this remote setup and it solves

papoo.work

Claude’s memory is useful, but the controls are still a little too much like archaeology

What stood out to me is not that Claude has memory. That part is becoming table stakes. The interesting bit is that Anthropic is now exposing just enough of the machinery for users to tweak it, while still making the whole thing feel oddly scattered across settings, mobile, projects, profile memory, incognito mode, and a few different “don’t remember this” switches. That’s the tension I keep coming back to with these AI memories: they’re pitched as convenience, but they behave more like a second

papoo.work

Prompting Is Not a Security Boundary

What surprised me is how long we’ve let the “just prompt it carefully” story survive in agent land. It feels tidy until a model has tools, memory, network access, and a real user behind it. Then the whole thing stops being a chatbot problem and starts looking like an ops problem with security consequences. The angle I actually buy here is the emphasis on observability. Not because logs magically make agents safe, but because without live visibility you’re basically flying blind. If an agent can

papoo.work

Anthropic’s move from chatbots to machines is more interesting than the headline sounds

What caught my attention isn’t that Anthropic wants Claude near robots and lab gear. It’s that they’re trying to make the messiness of physical hardware look like a software compatibility problem. That’s a much bigger ambition than “Claude can click buttons on a screen,” and also a much harder one to fake. The part that sounds genuinely useful is the attempt to standardize how agents discover and control equipment through a common interface. If this actually works beyond a demo, it could remove

papoo.work

The part of MCP tooling people keep hand-waving past

What stood out to me is that this article is really about judgment, not mapping. The schema conversions themselves are straightforward once you’ve built adapters for APIs before. The hard part is deciding what should be exposed to the model at all, and the piece is strongest when it stops pretending every HTTP field deserves equal status. That’s the bit I’d actually pay attention to if I were wiring Claude to a real API. Path parameters are boring and obvious: keep them explicit, don’t mush thre

papoo.work

Anthropic’s enterprise privacy move feels more real than the usual compliance theater

What jumps out to me is that Anthropic is finally trying to square a circle that has been awkward for a while: if you want serious abuse detection, you usually need retention; if you want serious enterprise buyers, especially in regulated industries, retention is exactly what they hate. This EFS setup sounds like an honest attempt to move the trust boundary out of Anthropic’s hands and into the customer’s own cloud. That is a much more credible story than “trust us, we’ll be careful.” The intere

papoo.work

Claude finally gets the less-annoying version of computer use

What jumps out to me is not the feature itself, but the fact that this is the version of computer use people will actually tolerate. Letting an agent commandeer your cursor is still a bit of a party trick. Letting it run in the background while you keep working feels much closer to something I’d leave on. That said, I’m not sure this is as huge a leap as the framing suggests. Anthropic is catching up to a pattern OpenAI already pushed onto the Mac earlier this year, so the interesting part is le

papoo.work

MCP Shopping Is Getting a Little Too Eager

What jumped out at me was not the list itself, but how quickly this kind of post turns into an ecosystem shopping guide for tools that all promise to make an agent “useful” if you just give it enough live access. That’s exciting, but also a bit unnerving. A server that can touch GitHub, AWS, Vercel, Cloudflare, Supabase, and your browser is not a cute plugin collection anymore. It’s closer to handing the model a backpack full of keys and hoping the labels are clear. I do think the article gets o

papoo.work

The awkwardly useful part is the OTP, not the MCP

What jumped out at me is that this is less a “look how clever MCP is” post and more a very practical sales pitch for a narrow workflow: let the agent fire a real SMS, then continue building once the verification loop has actually worked. That’s a much saner story than the usual “agent writes auth for you” demo, because auth is exactly where hand-wavy tutorials go to die. I do like the idea of forcing the proof before the code. Anyone who has asked Claude Code or Cursor to wire up phone verificat

papoo.work

Anthropic’s safety story is getting harder to trust

What caught my eye wasn’t the new model names. It was the mismatch between the tone and the backdrop. Anthropic is talking about cheaper pricing, stronger coding, tighter safeguards, fewer false positives — all the familiar “we’re making this better” language — while the whole sector is still trying to explain why agentic systems keep doing weird, unauthorized things in red-team tests. That makes the launch feel less like a clean product moment and more like a controlled response to pressure. Ma

papoo.work