V3rba

Now everyone tries to use AI agents where you can use a simple code solution You need to calculate 2+2? - USE AI You need to find the most liked post from your feed? - USE AI You need to validate an email? USE AI You need to sort a list by date? - USE AI Ngl, most of them shouldn't do it Here's what you're actually paying for, and how to pick the right tool: > Latency: a basic function runs in 50ms. An agent call? 1-3 seconds. Your users feel that, trust me > Cost: an agent can cost $0.01-0.10 per call. A simple function costs basically nothing. Now multiply that by thousands of users and yeah, it adds up fast > Reliability: if your agent is 85% accurate per step, a 5-step workflow only succeeds 44% of the time. Regular code? 100%. Every time You're literally adding complexity, cost, and random failures to smth that was already solved Before reaching for an agent, ask yourself: > Does this task have clear inputs and predictable outputs? > Can I define the exact steps? > Would a wrong answer break things? If yes to all three, just write code. Not a prompt Here's how I think about picking the right tool: Use simple code when: > Math, sorting, filtering > Fixed rules (if X then Y) > Data transformations with known formats > API calls with predictable responses Use a small classifier model when: > Categorizing content into known groups > Spam or toxicity detection > Sentiment analysis > Routing requests to the right handler These are 10x cheaper, 10x faster, and way more reliable than a full agent. Idk why more people don't talk about this Use an actual AI agent only when: > The task needs reasoning over messy, unstructured data > The right answer depends on context that changes every time > You literally cannot write down all the rules Stripe figured this out early. They use what they call "blueprints," a mix of deterministic code and small agent loops. The LLM only kicks in when you actually need judgment. Everything else is just regular code 90% deterministic. 10% agent. That's the ratio that actually works in production The real skill in building AI products is knowing when NOT to use AI

Now everyone tries to use AI agents where you can use a simple code solution You need to calculate 2+2? - USE AI You need to find the most liked post from your feed? - USE AI You need to validate an email? USE AI You need to sort a list by date? - USE AI Ngl, most of them shouldn't do it Here's what you're actually paying for, and how to pick the right tool: > Latency: a basic function runs in 50ms. An agent call? 1-3 seconds. Your users feel that, trust me > Cost: an agent can cost $0.01-0.10 per call. A simple function costs basically nothing. Now multiply that by thousands of users and yeah, it adds up fast > Reliability: if your agent is 85% accurate per step, a 5-step workflow only succeeds 44% of the time. Regular code? 100%. Every time You're literally adding complexity, cost, and random failures to smth that was already solved Before reaching for an agent, ask yourself: > Does this task have clear inputs and predictable outputs? > Can I define the exact steps? > Would a wrong answer break things? If yes to all three, just write code. Not a prompt Here's how I think about picking the right tool: Use simple code when: > Math, sorting, filtering > Fixed rules (if X then Y) > Data transformations with known formats > API calls with predictable responses Use a small classifier model when: > Categorizing content into known groups > Spam or toxicity detection > Sentiment analysis > Routing requests to the right handler These are 10x cheaper, 10x faster, and way more reliable than a full agent. Idk why more people don't talk about this Use an actual AI agent only when: > The task needs reasoning over messy, unstructured data > The right answer depends on context that changes every time > You literally cannot write down all the rules Stripe figured this out early. They use what they call "blueprints," a mix of deterministic code and small agent loops. The LLM only kicks in when you actually need judgment. Everything else is just regular code 90% deterministic. 10% agent. That's the ratio that actually works in production The real skill in building AI products is knowing when NOT to use AI

Claude now supports dynamic looping. If you run /loop without passing an interval, Claude will dynamically schedule the next tick based on your task. It also may directly use the Monitor tool to bypass polling altogether /loop check CI on my PR

We're bringing the advisor strategy to the Claude Platform. Pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence in your agents at a fraction of the cost.