Ariadne Dev Log, July 12, 2026

The agentic engineering experiment hits a wall

So I tried the whole loops thing. Plan, adversarial review, re-plan, repeat until it passes. Then build, test, fix, test, repeat. It works. It produces better output than just yoloing it. But DeepSeek V4 Pro still makes plenty of logical errors anyway. I've been doing a lot of manual review to catch stuff the loop misses, and it's going to slow down my internal expected release date.

I'm considering slotting in GLM 5.2 or an actual frontier model at specific stages. The stages where I'd normally need to step in myself. Limit how much they're actually used but still get the benefits. I don't know though. I still enjoy the manual tinkering. There's something satisfying about catching the agent's mistakes and fixing them yourself.

But I've confirmed one thing: fully agent-delegated software development is not viable with just DSV4 Pro. It gets you most of the way. Not all the way.

Splash screen, or: stuff I never had to think about before

Godot put things on the screen for me. You want a window? There's a window. You want a loading screen? Scene manager handles it. I never had to think about any of this.

Now I'm dealing with backend servers and localhost ports and browser windows, and this is all territory I have never bothered exploring. Before now, the setup was: run the backend server, hop onto a localhost port in my browser to interact with the frontend. Obviously temporary. But this stuff is going to be necessary for hosting cloud instances of the agent, accessing the UI over a self-hosted gateway, doing a mobile UI that works online. Plenty of things I have planned.

So I wanted to polish up the desktop app side. Get it to open a window that grabs the UI itself. Then I wanted a loading splash screen to show up while the server sets everything up in the background.

Here's the thing. The basic concept is:

Put splash screen in binary. Binary runs splash screen and throws the server up in the background. When it's ready, the binary fades from splash screen to the UI.

No matter how I explained this to DeepSeek, it did not want to move the splash screen out of the backend. Plain English. Different angles. Diagrams. It kept putting the splash screen in the server code. The binary just sat there waiting. The splash screen never left the backend.

Eventually I got it working through manual intervention. But this is exactly the kind of thing that makes me think about those frontier models for specific stages.

API calls and Godot comparisons

I'm working in unfamiliar territory now. API calls are simple enough. Godot signals with less abstraction, I can work with that. Learning the differences between the two is really easy thanks to LLMs. I think people underestimate how much that matters. You can just ask "what's the equivalent of a Godot signal in FastAPI" and it tells you. No tutorial hunting. No forum diving. It just maps it straight across.

So anyway, the unfamiliar territory I'm actually in now is the whole "proompting" thing. Prompt engineering. Agent orchestration. Context window management. Temperature and top-p and all the other sliders that I vaguely understand but couldn't explain to someone else. This is all brand new. I came from Godot. You make a scene. You add nodes. Things happen on screen. None of that prepared me for managing an LLM's attention span across a campaign interview and planning phase.

This is also where I need to learn and get outside knowledge the most. I can figure out API calls by comparing them to Godot signals. I cannot figure out why the agent keeps dumping the combat state from its context window by comparing it to anything I already know. There's no Godot equivalent. So for this stuff I'm actually reading documentation and watching talks and asking the LLM to explain concepts from scratch. It works. It tailors explanations to my skillset instantly. Normally that would require a tutorial from a very particular dev with a similar background. If you're coming from Godot and trying to learn FastAPI, there is no "FastAPI for Godot devs" tutorial. An LLM just does it.

Hermes gripes

One thing that's been annoying me: the manual context compression feature in Hermes has been broken for about a month. There are multiple PRs that fix it. I fixed it myself using one of those PRs, but I didn't do any git stuff for that code change, so a Hermes update regressed it. Now it's broken again. I don't really want to keep manually fixing it after every update.

What I'm working on now

Currently trying to get the comprehensive planning phase to correctly invoke after the campaign interview finishes. The interview runs, the builder agent collects all the answers, and then the plan generation is supposed to kick in. Getting the handoff right between those two stages has been finicky.