Last Week

Last week, I wanted to turn a promising video-production pilot into a repeatable workflow. I had better-looking Shokken tutorials coming out of Codex, but moving that workflow into Hermes was still taking a lot of hand-holding.

This week, I found a trap in asking one agent to teach another: the teaching can sound successful while losing the details that made the original work good.

This is Week 94 of building Shokken, my restaurant waitlist management tool. It is available on Google Play and the iOS App Store. The app is not the main subject of this update, though. This one is about the work behind the videos—and what happened when I tried to transfer it.

A Better Video Is a Collection of Small Decisions

I have been using OpenMontage with Codex to bring together the pieces of video production: storyboards, narration, keyframes, composition, encoding, and the smaller decisions inside each of those jobs.

The tutorial videos have improved considerably compared with what I was producing before. That is encouraging. If I can reduce the human effort required for each video, I can make useful material more consistently instead of treating every release as a separate manual production.

But the finished video hides how many decisions went into it. Where should an element sit? Which font should it use? Does the narration move at the right pace? How should one storyboard beat lead to the next? A neat final file does not make those choices obvious to the next agent.

Why I Still Want to Move to Hermes

There are two reasons I want less dependence on Codex: vendor dependency and customizability.

I do not want my whole workflow tied to one subscription, one company’s pricing, or my satisfaction with its latest product. I previously moved from Claude to OpenAI because I was unhappy with the experience I was getting. More recently, I have run into frustrating omissions and failures with OpenAI’s tools too.

I do not know how much of that comes from the model and how much comes from the harness—the surrounding application that gives the model tools, instructions, and a way to work. As a user, I receive them together. What I can judge is whether the result works for me.

The other issue is control. As far as I understand the Codex setup I am using, I cannot plug in my own memory provider in the way I want. I repeatedly find myself prompting it to check information we have already discussed, including details about the machine where I run Hermes.

In my own Hermes setup, I have configured a memory search every turn. The agent summarizes the request, a planner works out what to search for, and the memory system retrieves potentially relevant information. That does not guarantee perfect recall. It gives the system another chance to bring an important preference or piece of context into the conversation, and I have found that more enjoyable to work with.

The Transfer Was Lossy

I could not simply take the Codex memory I wanted and move it directly into Hermes. Instead, I asked the Codex agent to teach the Hermes agent the workflow.

I do not think this was deliberate sabotage or an agent refusing to train its replacement. It looked much more like a familiar teacher-and-student problem: knowing how to do something is not the same as conveying everything another person—or agent—needs to reproduce it.

After the first teaching session, I asked Hermes to make the video itself. The result was a disaster, but not because nothing had transferred. Quite a lot clearly had. The problem was the accumulation of small misses: alignment, font selection, layout, cadence, and other details that made the video much worse as a whole.

I was using the same OpenAI model in Hermes. Both harnesses were capable. In this attempt, the missing part was the teaching.

That distinction matters. Repeating the broad workflow more confidently would not necessarily expose the decisions that had been lost.

Transfer the Acceptance Criteria Too

The approach that has worked better so far was to ask Codex to examine its own process at a much finer level.

For each small task, what would count as an acceptable result? A storyboard needs a particular flow. A composed frame needs its caption and pictures placed correctly. Instead of transferring only instructions for making those things, I asked for the acceptance criteria and checks that would tell the receiving agent whether it had done them properly.

The aim was to give Hermes something more useful than “make a video like the last one.” It could compare its output with specific expectations and work out what still needed fixing.

That approach has a weakness too: the criteria are only as complete as the introspection that produces them. If the teaching agent never names an important quality, the receiving agent still has no reason to look for it.

I had to supply some of those expectations myself, including the alignment and visual consistency I associate with a professional-looking result. Credit where it is due, the agent had captured many of the things I wanted. I still needed roughly half a dozen to a dozen modifications to make the expectations closer to what I meant.

Supervision Helped, but It Was Not Free

After that, I asked Codex to supervise a Hermes production run: ask Hermes to produce the video, have Hermes check its own work, then examine the result and repeat the improvement loop.

That ran overnight, as I remember it, and nearly used up my weekly usage allowance. After two or three days of trying to hand over the skill, the Hermes result was within an acceptable range of quality.

Acceptable is not perfect. Some details were still lost. Most of the result was pretty good, and that is progress, but I do not want to call it a seamless or automatic migration.

This Is a Human-to-Agent Problem Too

The experience made me notice something I can miss in ordinary agent work. When I interact with an agent incrementally, I can see a problem as it appears and correct it in the next turn. That constant back-and-forth conceals how much teaching I am doing.

The gap becomes more obvious when I hand over a larger scope all at once: an existing codebase, a substantial project, or a production workflow with many small conventions.

My lesson is not that agents cannot teach each other. It is that knowledge transfer is work. For a large scope, I need to sit down with the agent, plan the transfer cooperatively, and work through it iteratively rather than expect a one-line request to preserve everything that matters.

Next Week

I will be traveling over the next couple of weeks, so I am curious how much I can get done with agents without spending all day in front of a computer supervising them.

I want to try a little vacation-and-work mode: take a break after walking, sit down with a drink, and use my phone to give the agents some work. I do not yet know how productive that will be. That is the next experiment.

If you are interested in the restaurant waitlist tool itself, visit Shokken. It is available on Google Play and the iOS App Store.