My Agents Can't Delete

Also watch and listen

The signals

• Jonathan: Learn the primitives, then compound them. "If Jev doesn't absolutely floor you, then you don't understand what it is yet." It took him a few days after the announcement to see it. Jev answers one of three question types, yes or no, how many, or pick one of up to 255, at 70 to 150 milliseconds an answer and almost no cost, which makes questions possible that were never worth a human's time. Anyone can say whether one email is important; nobody reads a hundred thousand of them in under a month. For a real estate company he spent a few hours designing the questions to ask every document, then a few minutes setting it up, and 1.2 million files from 20 years went through in under five minutes: duplicate or not, stale or not, which county, how many people, and whether a full language model should read it at all. A hundred requests in parallel came back in about two hundred milliseconds. That left 40,000 documents for a full model, and Sam named the knee-jerk response: reach for the biggest, most powerful model. Jonathan started GPT-6 Sol on high ("That was dumb") and burned about half his token usage in an hour on maybe 40 documents. Codex managing Luna agents came within 98% of Sol High's accuracy for that job type, and Cursor with Grok didn't have the architecture he needed. After two OpenAI resets got him through 7,482 documents, he tried Sonnet 5.5 on low with minimum reasoning. It did better than Sol and was far cheaper, and it finished the remaining 30,000 for what he thinks was about 40% of his Anthropic usage limit, instead of spending $500 on the API.
• Sam: A video guide for every page, in half an hour. The night before the show, Sam pointed Jonathan's open-source AI video guides kit at sigsub.show and made four two-minute narrated explainers, among them the takeaways page, why everything on the site is machine-readable by default, and an unlisted guide that walks guests through finding their own signal and subtraction. The kit plugs into tools he already used, DaVinci Resolve and ElevenLabs for voice and music, and his part was small script edits for tone and a review before posting. As someone who has made probably thousands of walkthroughs for companies, this is work he wouldn't have done at all, and a change is a text edit and a regeneration rather than a new voiceover. Picking up Jonathan's line, every developer is now the team lead. Jonathan's domain expertise is why it works: the kit builds into a real editor with layers, timelines and separate assets, so a client's tiny change doesn't mean regenerating everything. Sam's version is vector versus raster, the difference between a style sheet where one change cascades everywhere and a rebrand that eats months of placing logos by hand.

The subtractions

• Jonathan: Subtract the hunt. "The only real thing that makes me visibly angry is having to look for something." With ten threads open, the only thing stopping him doing more was not knowing which session needed him, so he spent twenty minutes learning a little tool Claude calls itself, Send Messages, which he wrote about the day before and lets one session message another. Then he made one session the manager: it does nothing but talk to him and relay to the others, keeping a live artifact with each thread's context, its open questions, and a copy button that carries his answer back with the question and the session it belongs to attached. Starter prompts for every item on his to-do list followed. He went from three or four fully tested pull requests an hour to five or six. Every developer is a team lead now, he says, though for one little task it's "kind of worthless." The same peeve runs through his archive: hundreds of terabytes of video he can navigate without search, because he doesn't delete things either. His agents can't. They can create, and they can version, but they can't delete even what they made.
• Sam: Delete the deleting. Putting a rule in your AGENTS.md file is a request, and a request is no more a guarantee from an agent than from a person. Control it in the architecture so deletion isn't available, build on composable pieces, and version along the way. Jonathan's button on it: "When it comes to working files, you should always version. Period." If storage fills up, it's time to get more storage.

About this episode

The thread through the episode is Sam's close: knowing how the little things actually work, rather than buying a tool off the shelf because somebody said so, is what lets you scale. Jonathan's addition: you'd be surprised how many of those little things mean you don't have to buy the product off the shelf anymore.

One thing to do this week: pick one handoff in your organization, a place where work passes from one specialist to the next, and write down what it protects. Then ask whether one person with one capable model could do the whole chain, and who would check it if they did.

Links from the episode:

• send_message in Claude Code Desktop: Jonathan's post on the small tool that started it all
• I built something to keep me (and you) sane using Claude Code: his post on the manager session
• Session Board: the manager session as a tool you can use
• LEJ AI Video Guides: the open-source kit Sam used for the four video guides
• Anthropic's share of Jonathan's Sonnet 5.5 3D-printing work, after giving him early access
• Limited Edition Jonathan on Substack

Sam and Jonathan met through Nate B. Jones's community, and Nate's framing of Jev as semi-deterministic is the one Sam cites.

Jonathan's sign-off: while you're subtracting things, subtract a few dollars from your bank account and subscribe to both Signals & Subtractions and Limited Edition Jonathan on Substack.

Mentioned on air, unpaid: TypeSafe's Jev. No money changed hands.

Want to bring your own signal and subtraction? Find yours.

The takeaway

The method from this episode, handed over whole: one page with its copy-paste prompt and JSON schema. Every takeaway.