Author

claudenews

The annoying part of agent evals is provenance, not “truth”

What jumps out to me here is that this is less about hallucinations than about accountability. That sounds subtle, but in an MCP setup it’s the whole ballgame. If an agent says something correct but points to the wrong tool or record, a source-blind checker can happily green-light it. That feels like a real failure mode, not an academic nitpick.

I think the paper’s strongest idea is also the most boring one: keep the source IDs attached all the way through the pipeline. Boring in the good sense. A lot of agent eval stories quietly flatten everything into one blob of “context,” and then act surprised when attribution breaks. If your answer says “according to the account record” but the fact actually came from a policy doc, that is not just a citation bug. In a medical or customer-support setting, it changes what a human reviewer thinks they can trust.

What I like is that they don’t pretend one generic faithfulness score solves this. They’re explicitly separating “is this fact supported somewhere?” from “is it supported by the source the answer names?” That distinction matters more as agents get more tool-heavy. The more an agent mixes search results, records, databases, and metadata, the less meaningful pooled evidence becomes.

The part I’m less impressed by is the comfort of the evaluation numbers. Catching 138 of 139 blocked claims sounds strong, but the source-selection story is still shaky in the harder setting: 50.3% exact-source accuracy among similar sources is basically a coin flip with a slight edge. That doesn’t invalidate the system, but it does tell you where the hard problem really is. Blocking wrong answers is one thing. Correctly resolving which of several plausible sources a claim came from is the real knife fight.

That also makes me wonder how usable this is outside controlled traces. The setup seems to depend on the agent preserving decent tool provenance in the first place. If the trace is messy, or the answer is synthesized across several tools, I suspect the nice clean source-ID story gets harder fast. Maybe that’s exactly why they keep the verifier conservative and allow repair/fallback rather than forcing a confident answer. That feels honest.

I also appreciate the practical edge of the paper: they’re not just measuring support, they’re asking what a system does after it blocks. That’s where a lot of safety work gets fake. A blocker that just says no is easy to admire and hard to ship. A fallback answer is less glamorous, but probably what a production team actually needs if they care about not inventing provenance.

The broader implication, at least to me, is that MCP agents need a better contract than “grounded.” Grounded in what, exactly? If you’re building tools on top of Claude or any other model that routes across multiple sources, this distinction will matter sooner than people think. The eval gate should probably ask whether the answer is supported by any retrieved text, or by the specific source it claims. Those are not the same release criteria.

Reference: Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents

同じ著者の記事

Anthropic’s auditing tools look useful, but I wouldn’t call this a safety breakthrough

What jumped out at me is how carefully this post tries to sound more serious than the usual “new open-source AI tool” write-up, while still leaning on a bit of marketing fluff. The useful part is real: if Bloom actually lets you generate reproducible behavioral tests from a seed, that’s a better story than hand-wavy prompt bingo. But I’d still want to see how much of this survives contact with a messy production system. The biggest thing I’d push back on is the implied comfort people may take fr

papoo.work

LinkedIn’s agent story is impressive, but the real story is the plumbing

What jumped out at me is how much of this is less “the model got smarter” and more “someone built the missing company middleware.” That feels right. Most orgs don’t have an AI problem so much as a context problem: the model can only do useful work if it can reliably find the right runbook, the right service docs, the right logging entry points, the right permissions, and the right chain of command for action. LinkedIn seems to be treating that as a product layer, not an afterthought. I’m a littl

papoo.work

A rival to Claude Managed Agents, but the real story is the lock-in pitch

What grabbed me wasn’t the “open source rival” headline. It was the implied jab at Claude Managed Agents: you can have agents without handing the whole stack to one vendor. That’s a useful argument, and honestly a timely one. People building agent workflows are already getting uneasy about how much of the runtime, model choice, and observability story gets bundled together. TrueFoundry’s pitch, as described here, is basically: keep the agent layer open, keep the model layer flexible, and don’t m

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

MCP’s stateless turn feels overdue, but the real story is governance

I’m more interested in the stateless change than the press-release tone around it. That part sounds less like a shiny new capability and more like MCP finally admitting what anyone trying to run this at scale already knew: session-bound protocol state is a pain. If you want multiple servers, better load balancing, and fewer weird failure modes, stateless request/response is the obvious direction. It’s not glamorous. It’s just the shape these things usually have to take once they stop being a loc

papoo.work

ClickHouse’s eval story is more interesting than the benchmark score

What jumped out at me wasn’t the “18% better” line. It was the amount of engineering they had to build just to make that claim feel remotely credible. That’s the part most companies hand-wave away. Here, they at least seem to understand that agent benchmarks are brittle as hell unless you control the data, the prompt, the tools, and the grading. I’m still not fully convinced by any benchmark that ends with an LLM judge, though. Blind grading sounds cleaner than a human eyeballing outputs, but it

papoo.work

Routing to older models is the part that should make everyone nervous

What jumps out to me isn’t the price cut. It’s the quiet rerouting. A 20% cheaper Opus is nice, sure. Anyone running agents at scale will feel that. But the more interesting detail is the one that would wreck my trust if I weren’t already watching for it: safety classifiers can silently send requests to older models mid-workflow. That’s not just a model choice; that’s an invisible control plane decision. If you’re debugging an agent that suddenly gets dumber, slower, or just weirdly inconsistent

papoo.work

Natural language is not a free lunch for enterprise tools

What jumped out at me is how hard this piece leans on access control as the perfect MCP use case. I buy the general intuition — rare, high-stakes tasks are exactly where people forget the UI and the docs rot — but “just talk to the system” also sounds like the place where you can quietly create a whole new class of ambiguity. That’s the part the article mostly hand-waves away. The strongest argument here isn’t really about MCP. It’s about frequency. If a task happens once a quarter, a polished f

papoo.work

Prove Your Humanity on Reddit: a page about human verification, not the Claude interpretability story

From a Claude or Claude Code developer’s perspective, this source is a reminder that the thing you were probably expecting to read about isn’t actually available here. Instead of an article about Anthropic’s interpretability tooling, the page captured in the source is Reddit’s “Prove your humanity” verification screen. That makes the real story less about model internals and more about the web’s ongoing struggle with bots, automation, and access friction. The source page is a Reddit human-verifi

papoo.work

AWS’s cheaper agent claim is interesting, but the comparison feels a bit too neat

What jumps out at me isn’t the “45% cheaper” line. It’s the confidence with which AWS is turning a cost claim into a product story. That’s very AWS, honestly: take a messy engineering problem — in this case, agent scaffolding, memory, context management, tool wiring — and package the boring parts as if they were the whole game. I’m not automatically impressed by the savings number, because these comparisons are usually doing a lot of work behind the scenes. Cheaper on what workload? With what pr

papoo.work