I Replaced Five AI Automations With One Weekly Job

I built five scheduled AI jobs to tell me which side project to work on.

Writing that sentence now, the problem feels fairly obvious.

I had been messing around with Hermes Agent and wanted a weekly project cleanup. Nothing ambitious. I have a bad habit of starting something, getting it 70% of the way there, and then getting distracted by a newer, shinier problem. I wanted a little nudge toward whichever project was actually worth finishing.

Somehow that became a collector for my notes, a project classifier, an urgency scorer, a weekly planner, and one final job that explained why the plan was smart.

It looked sophisticated. It had stages. It had logs. It had the general energy of something I might show in a demo and then quietly never use again.

The version I kept is one weekly job. It reads the active-project context, tells me which project is most worth continuing, and suggests what to pause. That's it.

The clean pipeline was the bug

Splitting work into small services is usually sensible. The trouble here was that I did not have five well-defined services. I had five model calls passing English to each other and hoping the meaning survived.

If the first job missed a note, the classifier worked from incomplete context. If the classifier called something urgent, the scoring job treated that word like data. By the time the planner produced a recommendation, several guesses had hardened into something that looked suspiciously like a fact.

The interfaces between the jobs were prose pretending to be types.

That was the part I had missed. Each job could do its local task reasonably well while making the whole system worse. A clean classification is not useful if the categories have nothing to do with the decision I need to make. A precise urgency score is not useful if the source material never contained a deadline.

I had automated the production of confidence.

One job, with an actual boundary

The replacement prompt is boring on purpose. It asks for one project to continue, the notes that support that choice, the projects I should pause, and any missing information that could change the answer. It also says not to invent deadlines or change project state.

That last part matters. I do not want a personal agent quietly deciding that a neglected project is dead because I did not mention it for two weeks. The agent prepares a decision. I still make it.

The input can be messy because summarizing messy input is where the model earns its keep. The output cannot be messy. I know where the brief appears, when I will read it, and what I am supposed to do afterward.

This is now the test I use for most of my AI automations: is the ambiguity in the input, or is the process itself still vague?

If I have inconsistent notes but know the exact brief I want, an LLM can be useful. If my request is “manage my career” or “keep me productive,” I have not described an automation. I have described a hope.

No amount of agent orchestration fixes that. Now the hope just has API access.

Most of my automations do not need an agent

I have started working upward from the least exciting tool.

If I only need to remember something at a certain time, I use a reminder. If the steps are fixed, I use a checklist. If the input is structured and the rules are stable, I would rather write a small script. One model call makes sense when I need to compress or classify language. An agent loop comes much later, when the job genuinely needs to inspect something, use a tool, observe the result, and retry within a boundary I understand.

That order has saved me from a lot of fun, unnecessary work.

“Fun” is not sarcastic there. Building the five-job version was fun. I like wiring systems together. I like watching one process hand a neat little artifact to another process. The problem is that the fun of building an automation and the value of owning it are barely related.

Every scheduled job becomes a tiny service I am now responsible for. It has credentials that expire, a prompt that may behave differently after a model update, output that needs somewhere to go, and a failure state I may not notice. Personal automations are especially sneaky because they rarely explode. They just keep producing increasingly stale little reports until I stop reading them.

My five-job system gave me more places to inspect without giving me a better way to judge the final answer. The logs could tell me which stage produced a sentence. They could not tell me whether working on that project was a good use of my Sunday.

The metric was whether I read it

I did not instrument this with a dashboard or calculate a percentage improvement. That would be continuing the bit.

The evidence is much simpler. The five-job version became a system I maintained. The single-job version became a brief I read.

I also know how to delete it. If I stop looking at the weekly output, the job goes away and nothing else breaks. That feels like an important property for software allowed to wander around my personal life.

I still think agents are useful. I just no longer measure an agent setup by how many autonomous steps I managed to squeeze into it. I care more about where the uncertainty enters, where the authority stops, and whether the output changes a real decision.

In this case, four-fifths of my architecture was mostly there to admire itself. Relatable, honestly.