Pratik Batha

Building Systems Memo (21k+ founders) | AI automation & agency workflows | Co-founder @Kynlyr | Built & sold a 120k newsletter

[Pratik Batha @pratik_batha]

@Charvikasliwal_ Hi

[Pratik Batha @pratik_batha]

I gave 6 AI agents a budget, a client list, and a live dashboard.

Then I watched them work.

Here's what nobody tells you about running an AI agency:

You don't lose money on the work.
You lose it on the agents you can't see.

One runaway loop can burn hundreds overnight.
No alert. No attribution. Just a bill.

So I built a control room with Claude Code:

→ Every agent run, cost and outcome in one live feed
→ Cost per client, per agent, per task
→ Auto-kill when an agent blows past its budget
→ Real margin per client, updated live

(Video = simulated demo data. The architecture is real.)

Want the exact Claude Code prompt + setup?

Free today.

Comment "AGENTS" and I'll DM it.

[Pratik Batha @pratik_batha]

I gave 6 AI agents a budget, a client list, and a live dashboard.

Then I watched them work.

Here's what nobody tells you about running an AI agency:

You don't lose money on the work.
You lose it on the agents you can't see.

One runaway loop can burn hundreds overnight.
No alert. No attribution. Just a bill.

So I built a control room with Claude Code:

→ Every agent run, cost and outcome in one live feed
→ Cost per client, per agent, per task
→ Auto-kill when an agent blows past its budget
→ Real margin per client, updated live

(Video = simulated demo data. The architecture is real.)

Want the exact Claude Code prompt + setup?

Free today.

Comment "AGENTS" and I'll DM it.

[Pratik Batha @pratik_batha]

A good agent demo proves it can work once.

It doesn’t prove it works when a tool fails or the data is thin.

Five checks before an AI system touches real work: Objective, Evidence, Verification, Recovery, Economics.

Issue #1 of Systems Memo:
https://t.co/Nrs6PMaQ9i

[Pratik Batha @pratik_batha]

Most AI agents don't fail because the model is too weak.

They fail because the system around the model is poorly designed.

Here are 7 things I'd check before giving an AI agent more autonomy 🧵

1/ Start with a narrow task.

If you can't explain the agent's job in one sentence, the scope is probably too broad.

2/ Give it reliable inputs.

Bad data doesn't become good data just because an LLM processes it.

3/ Define the tools it can use.

Every extra permission increases the number of things that can go wrong.

4/ Require evidence.

If the agent makes a research claim, it should identify the source supporting that claim.

5/ Design for failure.

Timeouts, invalid outputs, missing fields and API errors should have explicit handling.

6/ Add human approval.

High-impact actions shouldn't happen just because a model returned a confident answer.

7/ Measure the outcome.

A successful workflow execution isn't the same as saved hours, qualified leads, or revenue.

The goal isn't maximum autonomy.

It's a reliable system that does useful work.

I'm building Systems Memo around these practical questions: real workflows, honest tests, and the details behind what works.

What failure mode have you encountered most often with AI agents?

[Pratik Batha @pratik_batha]

AI agent reality check:

99 dealer records through n8n.
→ 0 duplicates
→ 86 website-research leads
→ 13 manual reviews

I tested AI on 5 sites. One timed out and stayed “insufficient evidence.”

Good automation knows when NOT to guess.

#n8n #AIAgents

[Pratik Batha @pratik_batha]

21,000+ subscribers.

Built around one idea: AI should do more than generate text. It should help businesses get work done.

I'm Pratik, founder of Systems Memo.

Now I'm going deeper into AI agents, business automation, and systems that help founders scale smarter.

I'm just getting started. 🧵

[Pratik Batha @pratik_batha]

21,000+ subscribers.

Built around one idea: AI should do more than generate text. It should help businesses get work done.

I'm Pratik, founder of Systems Memo.

Now I'm going deeper into AI agents, business automation, and systems that help founders scale smarter.

I'm just getting started. 🧵

[Pratik Batha @pratik_batha]

@KytroCrypto CA

[Pratik Batha @pratik_batha]

I got tired of doing LinkedIn outreach manually, so I built an AI workflow with n8n to automate it.

Lead finding, personalized messages, repetitive tasks.

Shared the whole breakdown here 👇
https://t.co/HtEOqQMsv3

[Pratik Batha @pratik_batha]

I got tired of doing LinkedIn outreach manually, so I built an AI workflow with n8n to automate it.

Lead finding, personalized messages, repetitive tasks.

Shared the whole breakdown here 👇
https://t.co/HtEOqQMsv3

[Pratik Batha @pratik_batha]

@stijnnoorman A newsletter is such a good way to go deeper than social media posts. You get to build a real connection with readers over time.

[Pratik Batha @pratik_batha]

@MillyTamati The compounding asset part is so real. Social posts disappear fast, but a good newsletter keeps building trust with readers over time.

[Pratik Batha @pratik_batha]

@heyakash_ai Google

[Pratik Batha @pratik_batha]

@Hey_AsiiF Turning a simple idea into a working system is getting easier with AI. The interesting part is whether the edge actually holds up over time.

[Pratik Batha @pratik_batha]

@vlatdd Yupp lets grow together!

[Pratik Batha @pratik_batha]

@m0agh Building in public, learning AI, and exploring useful tech tools. Always happy to connect with like-minded people 🤝

[Pratik Batha @pratik_batha]

@HanaFutureAI The interesting part isn’t Claude itself. It’s exploiting the gap between market prices and probabilities.

Speed matters, but execution and risk management matter just as much.

[Pratik Batha @pratik_batha]

@undefinedKi The separate agent grading the work is the part that stands out to me.

AI gets much more useful when it can verify its own progress instead of just generating output.

[Pratik Batha @pratik_batha]

@rafal_darlak Congrats on the 600! 🙌 I'm Pratik, building Systems Memo, a newsletter on AI workflows for founders: https://t.co/Q0PtS2vEZ8. Just started on here, would love to connect.