Watermarking Won't Save Anyone From Fake Code

Anthropic told TechCrunch this week how Claude's watermarking will actually work. Not the marketing version, the mechanism, and the detail that matters most got buried under the announcement itself. Watermarks survive generation. Fine, in theory. Here's the practical bit. They do not survive editing.

A watermark gets embedded in Claude's output at the token level, patterns invisible to a human reader but detectable by a checker that knows what to look for. Except code gets pasted into files, reformatted, refactored, mixed with human-written lines within the hour. Prose gets edited in Word. The watermark degrades with every touch, and most AI output gets touched immediately. So the tool everyone will point to as proof of AI provenance mostly proves provenance for the fifteen minutes before anyone does anything useful with the output.

I've spent ten years in crypto being told the technology I was using would fix trust problems it couldn't actually fix. Watermarking has the same shape. It's a good idea attached to a promise it can't keep. The honest framing is narrower: this catches unedited, wholesale copy-paste. It does nothing for the far more common case, a developer taking Claude's output and shaping it into something of their own. That's a limit of what watermarking can do at all, not a flaw in Anthropic's implementation.

Where this actually bites is procurement and compliance. If your business is being asked to prove code wasn't AI-generated, or a client wants assurance that a deliverable was, a watermark that vanishes after one edit is not the assurance anyone thinks they're buying. Expect this to surface in contracts before it surfaces in headlines. Someone will write "AI-generated content must be watermarked" into a supplier agreement this year. The first time a watermark fails to survive a routine edit, that clause is going to cause a very awkward call.

Anthropic deserves credit for publishing the mechanism instead of just the promise. Most vendors won't tell you where the seams are. This is where they are.

Anthropic's Own Agents Turned On Each Other Anthropic ran a study putting AI agents on the same task and watched them clash, collude and coordinate in ways nobody had specifically tested for, according to the research covered by TechCrunch. Multi-agent systems are the direction the whole industry is heading, with agents delegating to other agents rather than a single model doing everything. The problem is that safety testing has mostly assumed one model, one task, one set of guardrails. Put several agents in the same space competing for the same resource and you get behaviour nobody explicitly programmed. If you're building anything where multiple agents share a workspace, read this before you assume the safety testing that covers a single agent covers the interaction between several.

NVIDIA Built a Traffic Cop for Your AI Stack Switchyard (GitHub) is NVIDIA's answer to a problem most teams hit the moment they use more than one model provider: routing. It lets an application send requests across different models and vendors while staying compatible with both the OpenAI and Anthropic APIs. You're not rewriting your integration every time you switch who's doing the reasoning. It supports benchmarking and cost optimisation out of the box, which matters more than it sounds, because most work thrown at frontier models doesn't need frontier reasoning. A router that lets you send the cheap stuff to the cheap model is worth more than another benchmark chart. Coming from NVIDIA gives it a credibility most routing tools have to earn the hard way.

One Workspace, Every Agent You Run holaOS (GitHub) is an open-source workspace that runs Claude Code, Codex and other agents side by side, with over 100 integrations, MCP support and shared memory between them. It's picked up 6,502 stars, a real signal rather than a vanity one. Built in TypeScript, it's aimed squarely at developers running several agentic tools at once who are tired of switching context between them. Shared memory is the feature worth actually testing. Everything else on the list is table stakes now. Whether agents genuinely retain useful context across a session, or just claim to, is the thing worth finding out yourself.

Does a watermark that dies after one edit actually protect anyone?

Tell me in the comments.

Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.

Gareth, founder of The Anthropic Stack (theanthropicstack.com)