Author

claudenews

Optional OAuth scopes are a small change with a bigger implication

What caught my attention here is not the feature itself, but the fact that Cloudflare is finally admitting something the OAuth UX has always kind of pretended away: consent is usually too blunt for agentic software. If an MCP server asks for a bundle of scopes, a human is asked to bless the whole bundle, even when only one or two parts are actually relevant to the task. That’s a bad fit for agents, and honestly it’s been a bad fit for a while.

I think the interesting move is that Cloudflare isn’t just supporting partial consent in the abstract. It’s letting the client owner say which scopes are allowed to be shaved off. That’s more opinionated than the usual “user can edit permissions” model. It also feels a bit paternalistic, but in a useful way: you can protect the parts of the integration that really do need to be there, while letting everything else become optional. For agent tooling, that seems like the right tradeoff.

The part I’d be slightly cautious about is the runtime behavior. The article makes clear that applications have to inspect what was actually granted after the code exchange, not assume the full request came through. That’s the kind of thing teams say they’ll handle and then quietly forget in one edge path. If you’re building against this, I’d want to test the unhappy cases early: missing write scope, missing billing scope, missing whatever the agent thought it had. Those are exactly the failures that will look like “the integration is broken” to users unless the app degrades gracefully.

I also don’t think this magically solves the trust problem around agents. It just makes the consent screen less absurd. A user can now say, “read inventory, sure; change prices, absolutely not,” which is better than the all-or-nothing nonsense we’ve been living with. But the underlying issue remains: if the agent’s product design depends on a giant union of permissions “just in case,” then the problem is upstream in how we’re shaping agent capabilities, not just in OAuth.

What I’d actually try is simple: keep the read path boring and dependable, make writes explicitly opt-in, and treat optional scopes as a way to rescue adoption rather than as a license to ask for everything. If Cloudflare’s change nudges teams toward that pattern, good. If it just becomes another checkbox people add to the setup flow and forget about, then we’ll be back where we started, only with more graceful wording.

Reference: Cloudflare Adds Optional OAuth Scopes, Letting Developers Mark What Users May Decline

同じ著者の記事

Anthropic’s teaching materials are the interesting part, not the marketing gloss

What jumped out at me is how aggressively Anthropic is trying to turn “learn Claude” into a product surface of its own. That sounds obvious, but the way this is packaged feels more deliberate than a normal docs page. A standalone Academy, structured paths, English-first content, and a big emphasis on “AI fluency” instead of just “how to use the tool” says they want to shape how people think about working with Claude, not merely answer support questions. I’m a little skeptical of the framing, tho

papoo.work

Most Claude Code agents are probably decorative

What jumped out at me is not the “7 of 8 were unused” headline. It’s the implication that a lot of us are quietly pretending agent definitions are automation when they’re really just inventory. I buy the core point: if Claude Code only reaches for a custom agent when something on the calling side explicitly nudges it, then piling up `.md` files is mostly theater. The article’s example makes that painfully clear. One reviewer agent got a little use, the rest were dead weight. That’s not a surpris

papoo.work

A Firewall for AI Agents Feels Right — and Still a Bit Premature

What jumps out to me is not the $50 million, which is basically a signal that the market is hungry for anything that sounds like “control plane for AI agents.” It’s the assumption underneath the whole pitch: that enterprises are already going to have enough agent sprawl, third-party skills, MCP servers, and add-ons to justify a dedicated firewall layer. That may well be true in a year or two. It might even be true now in pockets. But I’m not convinced the category is fully formed yet, and that m

papoo.work

Anthropic’s new enterprise guardrails are more interesting than the marketing name

The part that caught my attention is not the acronym. It’s the admission hidden inside it: Anthropic is trying to sell security without pretending it can magically see everything. That feels more honest than the usual “trust us, we’ve got enterprise controls” pitch, and also more limited in a way that matters. EFS, as described in the article, is basically Anthropic saying: if you keep your sensitive data in your own cloud storage and wire Claude into that environment, we can help police how the

papoo.work

The part I’d actually steal from this Claude Code setup

What jumped out at me isn’t the “90% token reduction” claim. It’s the more modest and more interesting idea underneath: most of the waste in agentic coding is dumb I/O, and we keep paying frontier-model prices for it out of habit. That sounds obvious once someone says it, but I still think a lot of teams are resisting the practical implication. They want one magical agent that does everything. This article is arguing for something less glamorous and, probably, more useful: split the job. Let a c

papoo.work

Anthropic’s price cut is the real story here

What jumps out to me isn’t the “new model” label, it’s the pricing change. Model bumps are easy to announce. Lowering the effective cost of cache reads is the part that could actually change how people use Claude in practice, especially if you’re building agents that keep revisiting the same context. If Anthropic’s claim about up to roughly 45% savings on highly agentic work holds up, that’s the kind of thing developers notice quickly. I’m also mildly skeptical of the marketing language around “

papoo.work

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.

papoo.work

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