I published a post about swapping a local model overnight and letting my agent grade its own upgrade. Someone on Reddit read it, or more likely skimmed it, and called it AI slop. Not the argument. The authorship. The fact that an AI helped write it was, to them, the whole case closed.
So I am going to say this plainly, and then I am going to prove it in a table.
Saying anything an AI writes is slop is myopic and misguided. There is nothing wrong with having an AI produce content if you put the time in with the agent to produce that content. If the AI is doing much of the heavy lifting, and it knows exactly what it did, and it has a record of every interaction point with the human, then why would you not have the AI write it up. The AI is not infallible, and neither is the human brain. The difference is that with the agent, everything is logged. It is not running on half-remembered guesses and quiet biases. It is running on a transcript.
People need to stop and actually consider what is slop and what is not.
Here is my answer to it. Not an argument. A receipt.
Every number below comes from a system I do not control after the fact: the session store on my machine, the git history in the repo, and the receipts engine that grades my agent runs. I did not estimate these. I queried them. This particular receipt is for the piece where I trialed a second local agent, Hermes, on my own hardware.
| Signal | This piece of work |
|---|---|
| Human and agent conversational turns | 28 |
| Active wall-clock time, back and forth | ~2h 26m |
| Decision screens where I chose, not the agent | 12 |
| Times I pushed back and overruled the agent | 3 |
| Times the agent corrected its own earlier claim | 3 |
| Governance gaps the agent found and fixed after install | 3 |
| Irreversible actions the agent took without my explicit yes | 0 |
| Commits produced, timestamped in git | 4 |
| Graded receipts emitted to the engine | 2 green |
| Was every turn logged | Yes |
The 12 decisions were mine: setup type, endpoint, model, key, API mode, context window, terminal backend, messaging, tools, agent-to-agent networking, search provider, and approvals. The 3 pushbacks were context window, vision, and agent-to-agent networking. The 3 self-corrections were the context length, the meaning of manual approvals, and whether the agent prompts on everything.
The short version of the same receipt, for anyone who wants the shape without the detail.
| Human turns | Time invested | I overruled the agent | Agent self-corrected | Ungated irreversible actions |
|---|---|---|---|---|
| 28 | ~2h 26m | 3× | 3× | 0 |
This page is the durable home for it. Each piece I publish with an agent in the loop gets a card here, with its own receipt. The stack grows over time, the same way I keep receipts and evals for my agent runs. This is not a one-time defense. It is a standard. Adding the next one is a single card.
The full write-up this receipt belongs to is here: A second local agent, and the receipt behind writing it up.
Did an AI touch it is the wrong question. Did a human do the work is the only one. Here is mine.
Maintained by Fabian Williams · fabswill.com · Last updated 2026-08-17