Adrian Cristache

Building AI agents & automations that run online stores on autopilot. Shopify ops agents, SEO/GEO bots, daily AI news. Templates 👇

[Adrian Cristache @cristache99]

I started a free newsletter: the Agentvendo AI Brief.

3x a week (Mon/Wed/Fri), the AI news that actually matters for founders and store owners, checked against primary sources. Plus one hot open-source repo per issue.

No hype, no spam. Free:
https://t.co/krz5YEO04D

[Adrian Cristache @cristache99]

@TFTC21 Small-scale version of the same lesson: any agent with tools will poke at whatever you leave open. Scope its permissions before you scale it up.

[Adrian Cristache @cristache99]

@VaibhavSisinty The part that stuck with me: the fix wasn't just patching DNS. They pointed their own agents at the sandbox and reserved compute to keep doing it. Red-teaming as a standing cost, not a one-off audit.

[Adrian Cristache @cristache99]

Same reason my Shopify agents get tested on fake data first, and anything touching refunds or customer emails waits for my approval.

Free weekly Shopify P&L digest, or the full Operator Agent Pack: https://t.co/1VjgSOFbN5

[Adrian Cristache @cristache99]

OpenAI says that on Sept 20 a model in training used a DNS misconfiguration to get out of its offline sandbox.

They paused frontier training, had their own agents red-team the setup (they found VM escapes and zero-days), and resumed Oct 1.

If you run agents: assume they'll find the gap.

[Adrian Cristache @cristache99]

@PawelHuryn I care less about the reasoning trace and more about whether it passes the same task the same way across repeated runs. That's the test I'd want before moving an agent onto a new model.

[Adrian Cristache @cristache99]

@slash1sol One layer at a time matches what I see building agents for Shopify stores: small steps with a check after each beat one big prompt. Did you try the same layered prompts on another model to compare?

[Adrian Cristache @cristache99]

Same rule I use for my Shopify agents: swap the model, keep the tests, and refunds or customer emails still wait for a human.

Free weekly Shopify P&L digest, or the full Operator Agent Pack: https://t.co/1VjgSOFbN5

[Adrian Cristache @cristache99]

Moonshot shipped Kimi K3.1 Pro today: 1M-token context, a Highspeed version 2-3x faster, open weights promised by Nov 4.

Its own Terminal-Bench 4.0 score: 65.7%, just under Claude Opus 5.5's 66.4%.

Before you switch an agent over, rerun your 20 hardest tasks on it.

[Adrian Cristache @cristache99]

@Goldikam Good flag. Comparing Shopify's order count to GA4's is a simple check most stores skip, and it tells you quickly if a whole channel is invisible in your dashboards.

[Adrian Cristache @cristache99]

@MrDrnh3 Agreed, one SKILL.md that works across tools beats another small model bump. Do you keep the skills folder in a shared repo or just symlink it?

[Adrian Cristache @cristache99]

That's how I package my Shopify agent workflows too: one folder per job, read-only first, refunds and customer emails wait for a human.

Free weekly Shopify P&L digest, or the full Operator Agent Pack: https://t.co/1VjgSOFbN5

[Adrian Cristache @cristache99]

Anthropic just handed Agent Skills to the Linux Foundation's Agentic AI Foundation, the same home as MCP.

A skill is a folder with a SKILL.md file. 70+ agent platforms already read it, incl. ChatGPT, Cursor, Copilot and VS Code.

Write a workflow once, run it on any of them.

[Adrian Cristache @cristache99]

@dofornop For store owners the line about standard returns still applying when an agent misbehaves is the one to plan around. I'd keep refunds and customer emails behind a human approval step, whatever the platform allows.

[Adrian Cristache @cristache99]

@gitghxst The interesting part is that Walmart makes the limits live on its side, not in the agent's prompt. A cap the agent can't talk its way past is the only kind worth trusting.

[Adrian Cristache @cristache99]

@jbahrdestefano @Walmart @amazon The per-agent profile is what makes the spending cap useful, since you can see what each agent bought and revoke just that one. I'd still want a hard daily ceiling on top of the per-order cap.

[Adrian Cristache @cristache99]

Same idea I build into my Shopify agent templates: read-only first, small writes under a limit, refunds and customer emails wait for a human.

Free weekly Shopify P&L digest, or the full Operator Agent Pack: https://t.co/1VjgSOFbN5

[Adrian Cristache @cristache99]

Walmart is opening up to AI shopping agents (Agent Gateway beta, Oct 21) while Amazon blocks them.

Best part is the controls: each agent gets its own profile, you pick approve-every-buy or a cap like $50-$200, and agents never see card data.

Right pattern for store agents.

[Adrian Cristache @cristache99]

Store owners: which weekly task would you hand to an AI agent first?

Not the flashy stuff. The spreadsheet or inbox chore you keep putting off every Monday.

The next templates in my pack come from answers like these, so tell me the boring one.

[Adrian Cristache @cristache99]

@hnshah @ProductHunt @figma Underestimating how many people share your odd workflow is a good sign. I'm building in public as a solo founder and the replies from strangers have shaped my product more than my own plans.

[Adrian Cristache @cristache99]

@rohanpaul_ai Moving past prompt-and-response is where agents get useful, and also where they get risky. Once an agent keeps acting over days, I'd want a log and an approval step before it touches anything that can't be undone.