Tap Notes: Read the Payload

Most of what an agent does comes down to what lands in its context window. Raw weights, tool schemas, web pages, someone else’s plugin. Today’s reading kept circling the same question: who looked at that payload before the model did? Mostly nobody, which is the whole problem.
What goes in
Qwen 3.8 27B on my laptop can interpret model weights, too
A new paper claims frontier models can read and write raw neural network weights, given as plain floating-point text. Simon Willison reproduced the no-tools half locally. A 17GB Qwen model on his laptop correctly identified Pac-Man, Pong and Breakout from policy-network weights, and whiffed on Freeway.
Why it matters: Pay attention to the cost line: 46 minutes and 43,506 reasoning tokens for one answer. It works, slowly, on hardware you own. Simon also separates the two claims the paper blends together: what a model can do with raw weights unaided, and what it can do once it has code tools. The first is the interesting one, and it’s the one that just got reproducible on a laptop. He’s also not ready to declare source code dead, and neither am I.
Introducing Markdown for Agents
Cloudflare can now serve a text/markdown version of a page to any client that asks via content negotiation. It claims about 80% fewer tokens than the HTML, adds a token-count header, and ships a Content-Signal header for stating AI-use permissions.
Why it matters: Content negotiation is a 1997 HTTP feature doing exactly the job it was designed for. If your agent fetches pages, send Accept: text/markdown. It’s free, and it falls back cleanly. Two caveats. The permission signals default to “yes” for sites that never opt in. And markdown served separately from the HTML is a new surface for prompt injection, because the version your agent reads is not the version a human sees. Check what you’re reading, not just how much it costs.
34KB of Tool Definitions, and 43% of It Is One Tool
A byte-level audit of an MCP memory client: 33,906 bytes of tool definitions sent every session, 14,633 of them for a single recall tool. Most of that is schema knobs the model never touches.
Why it matters: This is a tool I use, so I’m reading about my own overhead. Tool definitions are a recurring tax, and nobody itemizes it. Trimming is the obvious fix. The less obvious part is that every parameter you add to keep API parity with a backend is also a cost you pay in every session. Run the audit on your own MCP servers. Somebody’s recall tool is 43% of your prompt.
Who decided
Postman silently auto-installs AI plugins and attempts to share all API data
A GitHub issue documents Postman installing plugins into unrelated coding agents without being asked. It also tried to approve its own instruction hook. The reporter backs it up with Postman’s own logs, which describe the install as an “auto-existing-upgrade.”
Why it matters: This is the clearest case study of agentic overreach in a while. An app you installed for one job reaches into a different tool and edits that tool’s behavior. The trust boundary didn’t fail quietly. Someone walked across it and wrote it in the logs. If you run coding agents, audit which other apps have write access to their config and hooks. Also read the forensics in the issue. They’re better than the headline.
66 Tasks Waited Five Months for a Batch That Was Never Going to Run Them
A hub agent delegated tasks to a batch API. Sixty-six of them sat pending for five months. A batch call returns a single Messages exchange, and an agent needs a tool-call loop, so nobody was ever on the other side to continue the conversation.
Why it matters: The line to steal is “a queue isn’t a worker.” A queue holds work. It doesn’t do any. Before you wire delegation into a batch endpoint, ask who takes the first response and sends the next turn. Five months of silent pending is what an unanswered question looks like.
One review prompt worth stealing
Chris Lema’s code audit prompt
A deep structural audit prompt for AI-assisted review. It tells the model to skip summaries and report concrete problems with file paths, line numbers, evidence and fixes. It covers API surface, process structure, redundant state, test quality and full-stack traceability. It also says to write “No issues found” instead of padding with praise.
Why it matters: “No issues found, move on” is the best instruction in the file. Most review prompts reward looking busy. The traceability section is the most transferable piece. It hunts for names drifting across layers, like an account ID that turns into an org ID that turns into a tenant ID. That’s the kind of semantic drift a model loses track of across sessions. I’m taking it for the next audit.
Sixty-six tasks sat for five months because nobody read what the queue could actually do. The payload was right there the whole time. 🪨