Your Agent Probably Needs Fewer Jobs

I have been experimenting with simple Hermes Agent automations lately. Nothing too dramatic, just small systems meant to smooth out some clutter in my digital life: reminders, recurring reviews, project check-ins, little prompts that help me notice what is slipping before it becomes a problem.

In theory, this is exactly the kind of thing agents should be good at. They can reduce remembering, compress scattered inputs, and turn vague digital noise into something you can actually act on. In practice, I overengineered a ton of shit first.

At one point I tried to build a personal project cleanup flow. The idea was simple: once a week, look at what I had been working on and tell me what deserved attention. Somehow that turned into one automation to collect notes, another to classify projects, another to score urgency, another to draft a weekly plan, and another to summarize why the plan made sense.

The version that actually helped was boring: one weekly check-in that listed active projects, called out the one most worth continuing, and suggested what to pause.

That is the trap I keep seeing with AI setups right now, and I fell into it too. A simple need turns into a small control room, and the whole thing starts feeling more impressive than useful.

Agents are useful. I am not arguing against them. But a lot of people are skipping the boring step: figure out the actual process first.

If the human workflow is unclear, adding an agent usually just automates the confusion. Now the confusion has logs.

Start Smaller Than You Want To

Before building an agent workflow, ask:

• Could this be a checklist?
• Could this be a reminder?
• Could this be a saved search?
• Could this be a template?
• Could this be one cron job with one clear output?

If yes, start there. That does not mean "do not use AI." It means do not summon a multi-agent mess when a calendar reminder would have done the job.

The goal is not to make the system feel powerful. The goal is to make the next action easier.

That is the line I would keep coming back to: does this make the next action easier?

If not, it is probably agent theater.

A Useful Agent Has A Small Job

The best agent tasks usually have a clear shape.

Bad prompt:

"Help me manage my career."

Better prompt:

"Every Friday morning, send me five roles worth looking at, why they fit, and one suggested next step for each."

Bad prompt:

"Monitor my personal projects."

Better prompt:

"Every Sunday night, list my active projects, identify the one most worth continuing, and suggest what to pause."

Bad prompt:

"Improve my content strategy."

Better prompt:

"Every Friday, draft one blog post idea based on my themes: user-owned tools, calm software, privacy, self-hosting, and humane design."

The better versions are narrow. They have a schedule. They produce a predictable output. They still leave judgment with the human. That is the part people forget.

Agents should not replace judgment. They should reduce the junk around judgment.

Do Not Add Another Inbox

Every automation creates a surface area. A channel. A notification. A dashboard. A database. A log. A thing to check. A thing to maintain. A thing that can quietly become stale.

This is where agent setups get sneaky. They promise to reduce work, but if you are not careful, they create another place where work collects.

Before adding a new channel or workflow, ask:

• Where will this output go?
• When will I look at it?
• What decision will it help me make?
• What happens if I ignore it for two weeks?

If you cannot answer those questions, you do not have an agent workflow yet. You have a content sprinkler, and nobody needs more sprinklers.

My Rule Of Thumb

Use an agent when it does at least one of these:

1. Reduces searching
2. Reduces remembering
3. Reduces repetitive formatting
4. Compresses messy inputs into a useful summary
5. Helps you make a specific decision
6. Produces a draft you will actually review

Avoid an agent when it mostly:

1. Creates more things to check
2. Requires frequent tuning
3. Produces output you skim but never use
4. Makes the system harder to understand
5. Exists because it seemed cool to build

That last one hurts, because "cool to build" is a very real force. Possibly the strongest force in software after "someone mentioned Kubernetes." But still: earn the complexity.

The Boring Version Is Probably Better

A good agent setup should feel almost boring. It should do one job, run at the right time, put the output where you already look, make the next action obvious, and be easy to turn off.

That is it. No orchestration diagram required. No glowing cyberpunk dashboard unless the dashboard is genuinely helping.

The point of agents is not to create a second nervous system for your life. The point is to remove friction from the one you already have.

So before building the elaborate setup, try the boring version first. One prompt. One schedule. One output. One decision.

If that works, then maybe add more. If it does not, the answer probably was not more agents. It was a clearer process.