The Workflow
Hermes read the SalesDaily funnel research I’d fed it, along with a stack of links, and came back with an actual strategy update: new positioning for the resource library, a website redesign, a shift in how I run social. That’s the easy part. Then I asked it to go execute the list.
Step 1 — Hand off the execution list: Strategy docs, website changes, the social shift. One list, in order.
Step 2 — Watch the loop: Hermes started handing tasks to the LLM behind it, ChatGPT 5.6-sol. The LLM would answer, Hermes would reply, and that back-and-forth ran for five or six minutes before Hermes quit with an execution error. No output, no partial progress I could point to.
Step 3 — Split it up: My fix for a stalled multi-agent conversation was more multi-agent. I broke the task list into separate Hermes conversations, one thread per item, figuring parallel threads would get further than one that kept looping.
Step 4 — Lose the thread, literally: Now I had five or six conversations running the same project and no single place that said what was actually done. I couldn’t tell which strategy sections were updated, which weren’t, or whether the website and social pieces had even started. Four days went by like that.
Step 5 — One AI, one conversation: This week, working the morning routine with Claude Code, I asked directly: is the strategy document I’m executing against the one Hermes actually updated, or some half-finished mix. One read-through against the original research file answered it. The part Hermes touched had landed clean. A different section, one Hermes never got to, was still describing a five-kit spine and a membership model I’d killed a month earlier. Fixed both, same conversation, no threads to reconcile.
I didn’t build anything new this week. I found out what four days of parallel AI threads actually cost me, and what one continuous conversation could do about it in about fifteen minutes.
What Broke
The failure wasn’t Hermes misreading the research. The read was good enough to become v2.16. It broke on execution: Hermes handing the actual work to ChatGPT 5.6-sol, one item at a time, waiting on a reply, replying back, looping like that for five or six minutes before quitting with an execution error and no clean readout, just a dead thread.
The fix that failed first was giving the loop more room: more threads, one per task item. That made the real problem worse, not better, because the failure was never “not enough parallel capacity.” It was that nothing was tracking state across the handoff. Splitting into five threads didn’t add tracking. It multiplied the number of places state could silently drop. The fix that actually worked was refusing the handoff pattern entirely: one AI, one continuous conversation, reading the finished output and the document it was supposed to update in the same context window.
An orchestrator handing work to a second AI and waiting for it to report back looks like delegation. It’s really two systems, each holding half the picture, trusting the other’s summary. When that trust breaks, nobody has the whole state. One AI checking one place with the whole picture ended it in about fifteen minutes.
Time Ledger
- Time saved: Zero. Nothing about this week was AI going faster than me.
- Time added: Roughly four days: the stalled Hermes/ChatGPT loop, then the failed fix of splitting into five parallel threads and losing track of what was actually finished.
- Net: Four days underwater, closed out in about fifteen minutes once one AI, in one conversation, actually checked the real state.
Fifteen minutes to fix costs less than the loop that broke it. The four days were the price of asking two AI systems to hand a job back and forth instead of just having one hold it.
The Prompt File
PROMPT FILE
You are a state-reconciliation auditor. I attempted a change with another
AI system or automation that may not have finished cleanly, and I'm not
sure whether the current version of a document reflects what was
actually decided.
Inputs:
[DOCUMENT_TO_VERIFY] — the file that should reflect the decision
[SOURCE_OF_THE_DECISION] — the original research, decision log, or
transcript where the decision was actually made
[WHAT_I_EXPECTED_TO_CHANGE] — a plain-language list of what I believe
should be different now
**Step 1 — Read both:** Read [DOCUMENT_TO_VERIFY] in full and
[SOURCE_OF_THE_DECISION] in full. Do not summarize either from memory
or prior context.
**Step 2 — List every decided item:** Pull every concrete decision out
of [SOURCE_OF_THE_DECISION] as a short bulleted list.
**Step 3 — Check each item against the document:** For each decided
item, state whether [DOCUMENT_TO_VERIFY] reflects it, partially
reflects it, or doesn't mention it. Quote the relevant line when one
exists.
**Step 4 — Flag anything untouched:** Separately, scan
[DOCUMENT_TO_VERIFY] for any section that looks like it predates
[SOURCE_OF_THE_DECISION] and was never updated, even if it wasn't part
of the original decision list. Name the section and quote the stale
line.
**Step 5 — Report, don't fix silently:** Return a short table: item,
status (landed / partial / missing), exact location. Do not edit the
document until I confirm which gaps to close.
Constraints: do not assume the document is current because it has a
recent version number. Do not skip Step 4 even if every item in the
decision list checks out clean. Never invent a decision that isn't
explicitly in [SOURCE_OF_THE_DECISION].Manager’s View
Ask whoever owns your AI rollout: if two different AI tools handed a task back and forth right now, would anyone actually know when it was done?
Most AI adoption plans measure which tools got deployed, not whether the handoff between them is legible to a human. When a task moves from an orchestrator to an execution model and back, “done” becomes whatever the last summary claimed, and nobody checks the actual document underneath it. A team running that setup can lose days to a stall that looks like busy activity, because from the outside there’s no way to trace it.
The fix isn’t a monitoring dashboard. It’s a cheap, specific check: before trusting that a chain finished, hand one capable AI both the finished output and the source it was supposed to update, in a single conversation, and ask it to reconcile the two directly. If that check isn’t fast and cheap to run, the handoff chain is too fragile to leave unsupervised.
Field Notes
- One day earlier, the same pattern showed up on a different project. Hermes/Codex tried to build an Apple Health tracking pipeline, a receiver service and a database, and didn’t finish. I cancelled it, handed the identical job to Cowork in one conversation, and had a working ingestion pipeline plus a weekly rollup by that same afternoon.
- The stale Section 1a had been describing a five-kit spine two months after the real catalog grew to eight, and a membership model a full month after I killed it. Nobody was executing off that section. Nobody had opened it either.
- This issue is itself four days later than it should be. Dispatch #10 drafts on Wednesday. The Wednesday it actually happened is the same one this story is about.