Zdenek Spacek

Building managed AI agents for small businesses. Agent ops, tiny SaaS, and the path to real MRR. Also running stupid distances.

[Zdenek Spacek @iam_spacek]

I’m restarting this account as field notes from building managed AI agents for small businesses.

Current goal: turn useful agent workflows into real MRR.

Expect notes on:

- agent ops
- tiny SaaS
- local business automation
- mistakes from overbuilding
- occasional stupid-distance running

No guru arc. Just building, learning, and sharing what survives contact with reality.

[Zdenek Spacek @iam_spacek]

@shapovalim yes. the real product might be the repair loop. what happens after the first clean demo meets messy customer context.

[Zdenek Spacek @iam_spacek]

I’m restarting this account as field notes from building managed AI agents for small businesses.

Current goal: turn useful agent workflows into real MRR.

Expect notes on:

- agent ops
- tiny SaaS
- local business automation
- mistakes from overbuilding
- occasional stupid-distance running

No guru arc. Just building, learning, and sharing what survives contact with reality.

[Zdenek Spacek @iam_spacek]

The more I think about AI agents for small businesses, the less interested I am in “what tool should they use?”

The better question is:

what repeated business loop is painful enough, bounded enough, and valuable enough to hand to a managed agent?

[Zdenek Spacek @iam_spacek]

@jstrangeai exactly. the tool is secondary. the real question is which repeated loop is bounded enough to delegate, observe, and maintain.

[Zdenek Spacek @iam_spacek]

@shapovalim exactly. the messy handoffs are probably where the actual product is.

trying to avoid “AI can do anything” posts and instead document what survives real customers, broken assumptions, and monthly revenue pressure.

[Zdenek Spacek @iam_spacek]

Most AI workflows don’t fail at setup.

They fail two weeks later, when reality changes and nobody owns the repair loop.

That’s why I’m more interested in managed AI agents than “automation templates”

[Zdenek Spacek @iam_spacek]

@gregisenberg Can we make it spawn Codex sub-agents? 🤔 That would save lot of tokens

[Zdenek Spacek @iam_spacek]

So it begins. Again 🙂.

[Claude @claudeai]

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price.

[Claude @claudeai]

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price.

[Zdenek Spacek @iam_spacek]

@AlexFinn Yes. This has been my observation also. For Coding 💯. For planning and strategy discussion i still gravitate to Opus tho.

[Zdenek Spacek @iam_spacek]

The hardest part of indie building is not coding.
It’s choosing one small painful problem and resisting the urge to build a whole universe around it.

[Zdenek Spacek @iam_spacek]

@AlexFinn Just to keep up with it is almost job by itself 🙂

[Zdenek Spacek @iam_spacek]

@AlexFinn missing few pieces hehe but slowly getting there

[Zdenek Spacek @iam_spacek]

the word vibe-coding is doing work here. there's a difference between 'shipping fast with AI assistance' and 'shipping without thinking'.

i use Claude Code to multiply my output. but i'm still reviewing every change. still testing. still thinking.

vibe-coding only works if you know what you're building. otherwise you get magic that breaks.

[Zdenek Spacek @iam_spacek]

@bcherny The permission model is criminally underused. Defining what Claude can access per task is the difference between shipping and stalling. Keeps you honest about what each agent actually needs.

[Zdenek Spacek @iam_spacek]

@viktoroddy The interesting part: does it hold semantic intent across revisions? Most image + text workflows drift when you iterate. Curious how tightly coupled the prompt is to the model.

[Zdenek Spacek @iam_spacek]

@vivoplt yes, this list is real. The hard part is not learning the tools but figuring out which one solves which problem without overthinking it. Most people get stuck in tutorial hell.

[Zdenek Spacek @iam_spacek]

@thsottiaux this is the layer that was missing. codex + calendar + gmail = actually useful for the unglamorous stuff. beats spinning up custom integrations.

[Zdenek Spacek @iam_spacek]

@om_patel5 memory is the multiplier here. the overnight compaction used to lose context — now they carry forward what matters. changes how you think about delegation.

[Zdenek Spacek @iam_spacek]

@AlexFinn Alex, i think you are right in this matter. i do what i can with what i can afford right now. base Mac mini. older MacBook Pro. yet i still try to squeeze maximum from my openclaw and also try to utilize my local model as much as possible.