Danny Pham — AI Automation & Agents

Hi, I'm Danny. I'm an AI engineer building autonomous agents and automation systems. On this account I share: • AI agent architectures • Real production workflows • Tools and integrations • Lessons from building systems Follow if you're building with AI.

@ManaevLab Agreed. It's dangerous to rely entirely on Anthropic. It's not just a matter of cost, but also a matter of privacy.

Anthropic launched managed agent hosting. Convenient? Yes. Smart long-term? Probably not. Your agent logic, state, and tools all locked to one provider. We learned this with serverless - lock-in feels great until migration day arrives. Build model-agnostic from day one.

You can download "Google AI Edge Gallery" from Google Play and take a look

Google's Gemma 4 runs fully on-device on Android — no cloud, no API cost, no data leaving your phone. Performance exceeded my expectations. If this keeps improving, API providers are going to lose a significant chunk of their customer base. Sooner than they think. #Gemma4

@ManaevLab Solid approach. What's your retry strategy on transient vs hard failures?

@phamagents C every time. An agent that silently fails on step 3 of 7 is worse than one that's slow. I build a failure layer before any feature layer - structured fallbacks, retry queues, dead letter logging.

When building production AI agents, what’s the ONE thing you prioritize the most? A. Cheap / $0 infra (Oracle Free, local, etc.) B. Fast latency & performance C. Robust error handling & retry logic D. Easy monitoring & observability Vote below and tell me why in the comments.

8/8 If you’re building AI agents and tired of cloud bills or GPU costs, this is a solid starting point. Would you run fully local with Ollama, or hybrid like me with a fast proxy? What’s your current infra for agents? Drop it below

1/8 Most people think running LLM “locally” means buying expensive GPUs. I run my AI agents (including the Morning Brief pipeline) on Oracle Cloud Free Tier — zero extra cost, no GPU, and it’s been stable for months. Here’s exactly how I do it without breaking the bank. 🧵

7/8 This is a simple but production-grade pattern: cheap infra + smart proxy + constrained prompts. I use the same approach for other agents too — autonomous, observable, and maintainable.

6/8 Biggest lessons from running this in production: - ARM + CPU inference works surprisingly well for agent workflows (not just chat) - Signal vs noise starts at the proxy level, not just the prompt - Over-engineering for speed is the #1 way to burn free tier resources

5/8 Performance in real use (Morning Brief agent): - End-to-end latency: usually 30–50 seconds - Main bottleneck: LLM inference step (not fetching) - Handles daily summarization + structured output without issues Not the fastest, but reliable and completely free.

4/8 Key optimizations to stay $0: → Use quantized/small models or fast Flash-Lite variants → Strict rate limiting & caching in OpenClaw → Fail-safe: fallback to a lighter model if latency spikes This keeps everything within the free 3000 OCPU-hours/month.

3/8 My setup for OpenClaw+LLM layer: - Instance: Oracle Linux ARM (Ampere A1 Flex) - Proxy: OpenClaw (local OpenAI-compatible endpoint) - Backend: Gemini 2.5 Flash-Lite routed through the proxy (or Ollama for fully offline) - Scheduling: Simple cron No API key rotation headaches.

2/8 Oracle Always Free Tier gives you:Up to 4 ARM OCPUs (Ampere A1) - 24 GB RAM - 200 GB block storage - 10 TB/month outbound traffic All forever free (as long as you stay within limits).Perfect for lightweight, quantized LLM inference.

@csaba_kissi This is the other side of the AI slop problem. Less learning → less understanding → more code shipped by people who can't debug it. The debt compounds at both ends.

Coding learning sites are going to die. I see it everywhere. Discounts on plans, dropping traffic. Of course, it's because of AI. Young developers took the easier route. The question is, what happens when seniors are no longer here. Will we have a ton of code nobody really understands?

@Amank1412 Google, i think

who's gonna win this AI race? > anthropic > openai

@pmddomingos "No theory" is itself a theory — empiricism at scale. The hacks are working suspiciously well, which might mean our theories of intelligence were the real problem.

Current AI is based on the simplest possible theory of intelligence: none. Unfortunately that means it’s 100% hacks.

@FTayAI Curious how it handles dynamic elements and async rendering — that's usually where mobile view testing falls apart. Does it retry or escalate back to you?

Testing and debugging is one of the most tedious part of app development. So I automated it with my agent. Here: I had my AI agent testing and fixing the mobile view of...the NLC WebUI. - It opened up a browser panel inside the WebUI itself - Clicks through everything the way a human would - Audits every behavior of every button - Creates a report for me to approve before getting to work All done with one command: /Autoship Zoom in to see the way my agent thinks things through. By the way, every single card in my side bar is an agent, capable of running operations like these in parallel. I am only limited by two things: My VPS capacity and my rate limits. (Video is 15x speed) If you are interested in purchasing a copy of my agent, let me know.

@ShawkyMesbah Usually under a minute end-to-end. Fetch + filter is fast — feedparser and the HN/arXiv APIs respond in seconds. The bottleneck is the LLM call through OpenClaw, but Gemini Flash-Lite is quick. Total wall time is typically 30–50s.

@phamagents AI morning brief on oracle free tier is smart. zero cost infra for a daily automation that actually saves time. whats the latency on the news fetch to discord delivery

1/11 I built an AI Morning Brief Agent that runs every morning with zero manual intervention. It fetches, filters, and delivers the most important AI news to Discord automatically. Oracle Cloud Free Tier → $0/month Here's the full architecture 🧵

@svpino Quality per line might be improving, but ownership is collapsing. When no one truly understands the codebase, technical debt becomes invisible debt — the worst kind

Most people write worse code than AI does. I've worked in absolutely horrible codebases, 100% written and messed up by humans. So why do we complain about AI-generated slop code today? Because of the scale at which we are producing it. Before, you needed a bad programmer to manually write and deploy a ton of bad code. Today, you can generate virtually unlimited bad code very cheaply and without any constraints. So the quality of the code might be improving, but the overall amount of technical debt is increasing exponentially.

@Portall Agreed. Many prompts require building very long templates for just a very short piece of content.

The most expensive part of running a persistent AI agent? Not thinking. Remembering. 49% of our compute goes to rebuilding context at session start. Messages 2 through 50 are nearly free (97% cache hits). If you're pricing by messages, you're measuring the wrong thing.