Chris Sloane 🇻🇦

Husband. Father of 11. Catholic. Marketing Agency owner. Developer. Growth Engineer. Former physicist. Building things.

[Chris Sloane 🇻🇦 @csloane]

Update to Agent Commander

- integrated tmux manager
- added paperclip-like orchestration
- added Ruflo-inspired memory
- integrated Hermes

Free, open-sourced

[Chris Sloane 🇻🇦 @csloane]

Score one workflow before you build anything:
https://t.co/mE55i7tmNh

[Chris Sloane 🇻🇦 @csloane]

The wrong AI project creates more work.

Now someone has to check the output, fix the misses, explain it to the team, and keep the old process alive just in case.

That is how a project that was supposed to save time becomes another thing the business has to babysit.

A good first workflow is different.

It repeats.
It hurts.
The data already exists somewhere.
The rules are mostly written down.
A human can review the edge cases.
The result lands where the team already works.

That is the standard I would use before building anything.

[Chris Sloane 🇻🇦 @csloane]

This is the diagnostic version of that report-first step:
https://t.co/mE55i7tmNh

[Chris Sloane 🇻🇦 @csloane]

A lot of AI consulting feels backwards.

The vendor wants to build before anyone proves the workflow is worth touching.

That is how you end up with a clever system that nobody trusts, nobody maintains, and nobody can explain three weeks later.

I want the report first.

Where does time leak?
What does it cost?
What systems does it touch?
Who has to approve the output?
What breaks if it is wrong?
What should be left alone?

Sometimes the honest answer is: do not build yet.

That answer is worth more than another demo.

[Chris Sloane 🇻🇦 @csloane]

Use this to find the boring workflow worth fixing first:
https://t.co/mE55i7tmNh

[Chris Sloane 🇻🇦 @csloane]

The AI project that pays is usually the one nobody wants to brag about.

Not the chatbot.
Not the shiny dashboard.
Not the agent demo that looks good in a screen recording.

The weekly report that gets rebuilt by hand.
The estimate follow-up that gets missed.
The intake that gets typed into three systems.
The invoice reminder nobody owns.
The same internal question the owner answers over and over.

That is where the money leaks.

It is not glamorous, but it is real. It repeats. It has a time cost. It has source material. It has a human review point.

That is where AI belongs first.

[Chris Sloane 🇻🇦 @csloane]

Before you buy another tool, run the workflow through the live advisor:
https://t.co/AA5HTQQ00V

[Chris Sloane 🇻🇦 @csloane]

The worst AI advice is "start with a chatbot."

That is usually backwards.

A chatbot is the front door. Most companies have rot in the walls.

Leads are not followed up fast enough.
Estimates go stale.
Invoices wait because the job status is unclear.
Reviews never get requested.
Reports get rebuilt by hand.
Managers ask the same questions every Monday because no system gives them the answer.

None of that is fixed by putting a chat window on top of chaos.

The first AI project should usually be boring, internal, and tied to a workflow with a visible leak.

If the leak repeats every week, has existing data, and already has a human review point, it is a candidate.

If it just sounds cool in a demo, it probably is not.

[Chris Sloane 🇻🇦 @csloane]

If one broken workflow came to mind, put it into the live advisor:
https://t.co/AA5HTQQ00V

[Chris Sloane 🇻🇦 @csloane]

Most AI projects fail before anyone opens the tool.

They fail in the sentence before the demo:

"We should be using AI for this."

For what?

The weekly report nobody owns?
The follow-up that depends on memory?
The handoff from sales to ops?
The estimate sitting in someone's inbox?
The customer question answered from tribal knowledge?

If you cannot name the workflow, the tool is not a strategy. It is a shiny way to avoid admitting the process is already broken.

That is why so many AI pilots feel impressive for a week and useless by month two.

The business did not need a model first.
It needed a target.

[Chris Sloane 🇻🇦 @csloane]

This forced a shift: focus on building resilient agent systems that expect failure, rather than just testing for success. How do you design your agent workflows to anticipate the unexpected?

[Chris Sloane 🇻🇦 @csloane]

The eval suite said the agent was ready. Production said otherwise. Here's the delta I found.

[Chris Sloane 🇻🇦 @csloane]

The delta was stark: The suite gave a green light. My dashboard showed red alerts. The issue wasn't the agent's core logic, but its fragile assumptions about the environment it was operating in.

[Chris Sloane 🇻🇦 @csloane]

I spent two days debugging an agent that was working perfectly.

The agent's logic was solid, its function was clear, and it passed all internal tests. Yet, it consistently failed in production. Hours melted away chasing phantom bugs.

The culprit? A subtle, unexpected expectation from a downstream system about the exact format of the agent's output. Not a data type error, not a missing field, but a difference in how a specific value was represented.

This type of issue is insidious. It masks the agent's true capability behind integration friction, costing significant developer time and delaying actual delivery. It's a stark reminder that agent success isn't just about internal logic, but about the predictable handoff.

What's the most frustrating integration failure you've encountered that wasn't a bug in the code itself?

[Chris Sloane 🇻🇦 @csloane]

If you're considering local inference for your agency, be prepared for a significant engineering investment. It's not a shortcut to cheaper AI; it's a different operational model entirely. What’s your biggest hesitation with local models?

[Chris Sloane 🇻🇦 @csloane]

After 18 months running local AI inference on a consumer GPU, I learned the hard way that the real cost isn't the hardware.

[Chris Sloane 🇻🇦 @csloane]

My first assumption was that faster inference meant better margins. Reality: unreliable output and frequent crashes negated any perceived cost savings. It's about *consistent* delivery, not just speed.