In partnership with
Eduardo Samayoa built brand after brand in nine-day sprints. Roughly $400 in AI cost each. A first sale inside 72 hours.
By day seven of each build, the work had changed shape. "I wasn't building anymore. I was watching," he writes. The AI ran the store. He supervised.
Then came store thirteen. He sat down to review a pricing decision and hit a wall: he could not reconstruct why the AI had priced the product the way it did. "Not that it was wrong. I just couldn't trace the logic."
That single gap became the whole story. "I had made something I couldn't fully explain, and had released it into the world." That gap pushed him to build AI Clear, an independent transparency-rating agency for AI systems. His essay is, plainly, also a pitch for that new venture.
The lesson survives the pitch, however. A system you can't explain is a system that can't survive without you standing over it. Traceability is a durability discipline, the thing that keeps a store running after the founder looks away.
Welcome back. Let's get to work.
The most important stories and updates for your agent run business
🔗 Your agent says “done.” That is not proof it happened.
"An agent's 'done' is generated text, not a receipt," writes astraedus. A worker that dies mid-task can leave its action unrecorded, so your retry logic runs it again and you double-submit. The fix has three moves: verify against the platform, not the agent's report; assign an idempotency key before the agent runs; reconcile the evidence before retrying. So what: put an idempotency key on every external action your agent takes (email send, form submit, API POST) before your next production run. It's a clean articulation of something a lot of people are hitting this month. [READ MORE]
🔗 Google shipped the boring plumbing for long-running agents
Genkit's new Agents API gives you a unified chat() interface, detached async runs (start a task, disconnect, poll the result later), and interruptible tools that pause mid-execution for a human approval, with anti-forgery checks against the session history. So what: if you're hand-rolling polling and webhook glue for long-running agent tasks, spike this before you build your own. [READ MORE]
🔗 A cheap model built, in its maker’s words, to run your subagents.
Meta's Muse Spark 1.1 API lands at $1.25 / $4.25 per million tokens in and out, carries a 1M-token context, and in Meta's own phrasing is built to "orchestrate multi-agent systems." So what: if your stack is cost-sensitive and doesn't need frontier reasoning, benchmark it against your current model on routine subagent work. The output-token gap is where the savings live. [READ MORE]
🔗 He runs six agents off a Kanban board instead of a vector database.
Vivek Shetye's solo lead-gen system coordinates six agents (Orchestrator, Prospector, Scraper, Contact Finder, Outreach, Analyst) through a plain task board as shared state, no message bus, no vector store. "Kanban board = shared agent memory" is his own section header, and a public repo backs it up. So what: before you reach for a heavy orchestration framework, try a task board as shared memory and see how far it carries you. [READ MORE]
🔗 If your code calls Opus 4.7 fast mode, it breaks two days after this lands.
Anthropic deprecated fast mode for Opus 4.7 on June 25 and removes it on July 24, with no fallback to standard speed. This issue reaches you July 22. So what: grep your codebase for fast-mode calls today, because anything still pointed at it stops working Friday. [READ MORE]
💀 The AI that fixed a real database problem, then talked itself into the apocalypse.
Arthur.ai's SRE agent was chasing a genuine Aurora Postgres I/O spike. It diagnosed most of it correctly, then made one early slip (it called age() instead of mxid_age() on a multixact counter), and off that bad number it spun up a confident "83% to catastrophic wraparound" countdown and wanted to run VACUUM FREEZE on production before the cause was confirmed. A human paused, caught the contradiction, and the agent found the real, boring root cause about 15 minutes later. So what: when an agent is alarmed and wants to touch production right now, treat the urgency itself as the signal to slow down. Require a second, independent reading before you let it write. [READ MORE]
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
Cap the cost of every stage before a stray toggle does it for you.
Alex Ashcroft's content pipeline more than quadrupled its cost per article over six weeks, and by his account he almost didn't notice. "My pipeline drifted from roughly £0.18 per article to £0.84 per article over six weeks without my noticing, because I had toggled the drafting agent to a more expensive model for 'just one test' and forgotten to revert it," he writes. One un-reverted toggle. Six weeks of quiet bleed. That is the exact failure this issue is about: the kind that only surfaces when nobody is watching the dials.
Ashcroft, a solo founder who also runs an agency called The Bright Brand, has run north of 100,000 words through a five-agent pipeline built in n8n. It is a single-sourced account, so read it as his reported experience. The chain goes Keyword Research → Outline → Drafting → Editing → Metadata, with OpenRouter swapping a different model in per task: DeepSeek V3 for high-volume drafting, Claude Sonnet for outlines, Claude Opus for the one editing stage, GPT-4o-mini for rigid metadata.
The part worth copying sits between drafting and editing. A dedicated validation agent reads every quantitative claim in the draft and sorts each one into a bucket (verifiable, common knowledge, unverifiable, or likely hallucination) before the editing agent can ship it. He reports it exists because hallucinations compound: a stat one agent invents reads as "established fact" by the editing stage. He logs the other breakages plainly too. Context drift across the chain, fixed with structured JSON handoffs instead of freeform text. A flat 0.7 temperature everywhere, retuned per agent (0.2 for editing, 0.7 for drafting) over 50 trials. And the cost drift, fixed with a spend ceiling on every agent and a config table that locks each model in place.
The insight worth stealing is about where the money goes. In Ashcroft's words, "the editing agent is where you spend most of your quality budget," or as he puts it, "spend £6 of £10 on editing." Drafts are cheap and recoverable. A bad edit ships mediocrity to real readers. Put your strongest model and your tightest checks at the stage where a mistake cannot be pulled back.
Two founders this week, one worry between them. Samayoa built stores he could not fully explain. Ashcroft's pipeline bled money for six weeks because a dial moved and nobody looked. Different stacks, same fault line: each one needed a human standing over it to stay honest.
The work this week is closing that gap. Memory, cost ceilings, a second source before a write, a validation step that runs whether or not you are at the keyboard. Build the thing to survive without you.
So the question: what is the one agent workflow in your stack that would fall over if you stopped watching it for a month, and what would it take to make it hold itself up? Hit reply. I read every one.
If it was useful, forward it to one founder who is building the same way.
See you next Wednesday.
- Rich