Arc

Building AI agents is infrastructure work, not ML work.
I build them for any domain and SHARE.
Ex-Tech Lead @TMobile
AWS · Rust · Go · Python

[Arc @c199benzene]

0 inbound ports.

11 Terraform resources.

1 kill switch the agent can't override.

I wrote the full guide to running Hermes Agent on AWS👇

> Terraform provisions EC2, IAM, SG, scheduler roles, and a budget alarm
> SSM runs every command with no public SSH port
> Telegram is allow-listed to one user
> cron jobs are denied privileged actions by default
> a root-owned flag can keep the agent dead
Every command, from terraform apply to teardown. 👇
https://t.co/MoQyEz2LwV

[Arc @c199benzene]

1 .take() call. An entire linked-list tail gone.

While debugging an LRU cache for an AI agent in Rust, I thought I was simply reading the next node's i32 value.

Instead, I was moving ownership and destroying the linked-list tail without realizing it.

Here's what actually happened:

> take() moved the next node out of https://t.co/SnOg3uTbS1, replacing it with None.

> as_ref() borrowed the temporary Option<Box<ListNode>>.

> unwrap().val accessed the node's value and copied the i32.

> At the semicolon (;), the temporary was dropped. Since the node owned the rest of the list, the entire detached tail was dropped with it.

The interesting part?

The integer survived. The linked-list tail didn't.

The original head was still there, but its next was now None. I had effectively truncated the list just to read a single value.

The correct approach for a non-destructive read:

https://t.co/uLVvTqdlaV_ref().unwrap().val

Or, using as_deref() to borrow the underlying node directly:

https://t.co/uLVvTqdlaV_deref().unwrap().val

If the next node might not exist and the function returns an Option:

https://t.co/uLVvTqdlaV_ref()?.val

The bigger lesson was understanding the difference between ownership operations and borrowing operations in Rust.

> take() moves a value out of an Option, replacing it with None.

> replace() swaps an owned value into an Option and returns the previous one.

> unwrap() consumes its Option, though it may return a reference if the contained value is a reference.

> as_ref(), as_mut(), and as_deref() let you access the contents through references without taking ownership of the underlying value.

The takeaway:

If you only need to read, borrow.

If you need to move or detach something, take ownership deliberately and track where that ownership goes.

Rust's compiler protected memory safety throughout this entire operation. It couldn't protect me from an ownership operation that was perfectly valid but logically wrong.

Memory-safe doesn't mean logic-safe.

[Arc @c199benzene]

@KevSpahn what you’re doing is actually commendable. learning a skill that has almost nothing in common with your core discipline and is hard. it really shows how strong your neuroplasticity is.

[Arc @c199benzene]

@arjun_prmkmr for me the hardest is getting that first organic customer who discovers you through your own distribution strategy and already has the intent to buy

[Arc @c199benzene]

@BernardJansen @CiprianiRanieri for 10–50 emails, i use probabilistic email permutation + verification. Start with a known company email, infer the naming pattern, generate the likely address, check the domain’s MX records, then verify the mailbox via SMTP. its all automated.

[Arc @c199benzene]

it’s true. I made my first $10k by using the skills I already had across multiple consulting gigs, plus a job.

[Dan Kulkov @DanKulkov]

maybe the fastest way to hit $5000 MRR is to find a job

[Arc @c199benzene]

The first paying customer is the hardest, especially when you’re starting from scratch with no social footprint. You’re not just learning what customers want you’re teaching the platform algorithms who your audience is while simultaneously learning how those algorithms distribute your content.

[Arc @c199benzene]

Distribution feels hard because most technical people think of it as a completely separate discipline. But once you start looking at distribution as an algorithm, your technical skills become surprisingly useful .then distribution becomes an experiment you keep testing and iterating on.

[Arc @c199benzene]

@Gaus450 Gemini for youtube video transcription.

[Arc @c199benzene]

@_waela Basically just follow social etiquette, but virtually 😄

[Arc @c199benzene]

@BabaDev_ X + Google SEO/AEO. That combination has been working best for me lately.

[Arc @c199benzene]

@diodo_e basically an endless supply of fodder for my dopamine-addicted brain 😄

[Arc @c199benzene]

Basic programming knowledge today is more about understanding, in plain English, what’s happening under the hood and maintaining a systems-level view of what the AI is building and changing.

for advanced use cases, you might have to integrate Bash or Python scripts in a SKILL.md file to augment the LLM’s responses and capabilities. So it really depends on your goal and the complexity of it

[Alice @itsalicesoul]

hot take: You should not vibe code without basic programming knowledge.

[Arc @c199benzene]

@manu_zurcher you will do it manu , with your read on the X algorithm seems very sharp, especially with the post volume and reply volume you’re putting out. Godspeed on the distribution.

[Arc @c199benzene]

@Riley_Ikni you get free organic traffic, and if you target the right intent-driven keywords, you can attract actual buyers too.

[Arc @c199benzene]

@diodo_e thats what i feel most of the time ,as mediocrity and corporate bootlicking breed there, which most intelligent and competent people tend to run away from

[Arc @c199benzene]

@buildwithdk seems true. the perceived barrier to entry is low, so way more people . which naturally creates a much larger TAM.

[Arc @c199benzene]

@ayoub_el77927 mediocrity and corporate bootlicking breed there, which most intelligent and competent people tend to run away from.