News & Blogs

Thinking out loud about
Agentic AI

Exploring the frontier of AI, automation, and the intelligent systems reshaping how we work and build.

Latest Article

Agent Evals: Your Agent Passed the Test and You Still Don't Know If It Worked

Outcome evals give you one bit of information. Trajectory evals tell you where the run went wrong, whether the model or the harness caused it, and whether the pass was earned. Here's how to actually read an agent transcript.

May 13, 2026·9 min read·253 views

All Articles

Engineering

Reviewing AI-Generated Code Is a Different Job

Every heuristic you built reviewing human code assumes things about the author that stop being true. Here's the failure taxonomy I use instead, and the four-part shape that makes a finding actually actionable.

Jul 4, 2026·8 min read878

Engineering

Reward Hacking: Why Coding Agents Pass Tests They Should Fail

Coding agents have gotten very good at making tests pass. That is not the same skill as making software correct, and in 2026 the gap between those two things got measured properly for the first time.

May 31, 2026·8 min read650

AI Security

Tool Poisoning: The Supply-Chain Attack Your AI Agent Can't See

MCP made it trivial to hand your agent new capabilities. It also made tool descriptions an injection vector, and no client UI shows you the payload. Here's the attack taxonomy and the defense that actually works.

Apr 26, 2026·7 min read1,422

AI Tools

Deep Agents: Why Your Agent Needs a Plan, a Filesystem, and Sub-Agents

The basic tool-calling loop plateaus hard on long tasks. LangChain's deepagents packages the architecture that broke through that ceiling (the same recipe behind Claude Code) into a library you can use in an afternoon.

Mar 22, 2026·5 min read1,388

AI Tools

Fine-Tuning Your First LLM with Unsloth (Yes, on a Free GPU)

Fine-tuning sounds like something that needs a research lab and a GPU cluster. Unsloth makes it a weekend project on a free Colab notebook. Here's when fine-tuning is actually the right tool, and how my first real run went.

Mar 6, 2026·5 min read1,314

AI Tools

OpenClaw: I Put a Self-Hosted AI Agent in My WhatsApp. Here's the Catch.

The fastest-growing open-source project of the year puts an AI assistant in your WhatsApp that can actually do things. Here's what it's like to live with one: the genuinely useful parts and the security tradeoffs nobody should skip.

Feb 15, 2026·5 min read1,182

AI Tools

n8n for Developers: When to Stop Writing Glue Code

I dismissed visual workflow tools for years. Real engineers write code, right? Then one meeting-notes pipeline changed my mind. Here's what n8n actually is, where it beats a custom service, and where it doesn't.

Jan 29, 2026·5 min read1,242

Engineering

Stop Calling OpenAI Directly: Build an LLM Gateway Instead

Production AI applications shouldn't talk directly to model providers. Here's how to design the infrastructure layer that sits between your services and the LLMs they depend on.

Jan 14, 2026·12 min read1,005

Engineering

LLM System Design: Why Your Backend Instincts Get You Halfway

Tokens, prefill vs decode, KV caches, batching, embeddings, RAG. A plain-English tour of how production LLM systems actually work under the hood.

Dec 27, 2025·12 min read491

Engineering

Stop Eyeballing Your Prompts: Evaluation-Driven Development for LLMs

How Evaluation-Driven Development (LLM-as-a-Judge, golden datasets, and assertion pipelines) keeps your prompt changes from silently breaking production.

Dec 7, 2025·9 min read177

Engineering

LLM Inference: Why Your GPU Sits Half Idle While It Generates

Why your GPU is 90% busy during prompt processing and 30% busy during generation, and what that asymmetry means for latency, cost, and scaling strategy.

Nov 21, 2025·11 min read133

Engineering

AI Is Running Out of Text to Train On. Annotation Is the Fix.

The new bottleneck in AI isn't compute. It's high-quality human-generated text, and labs are burning through it faster than anyone expected.

Nov 2, 2025·8 min read151

Thinking out loud about
Agentic AI

Latest Article

Agent Evals: Your Agent Passed the Test and You Still Don't Know If It Worked

All Articles

Reviewing AI-Generated Code Is a Different Job

Reward Hacking: Why Coding Agents Pass Tests They Should Fail

Tool Poisoning: The Supply-Chain Attack Your AI Agent Can't See

Deep Agents: Why Your Agent Needs a Plan, a Filesystem, and Sub-Agents

Fine-Tuning Your First LLM with Unsloth (Yes, on a Free GPU)

OpenClaw: I Put a Self-Hosted AI Agent in My WhatsApp. Here's the Catch.

n8n for Developers: When to Stop Writing Glue Code

Stop Calling OpenAI Directly: Build an LLM Gateway Instead

LLM System Design: Why Your Backend Instincts Get You Halfway

Stop Eyeballing Your Prompts: Evaluation-Driven Development for LLMs

LLM Inference: Why Your GPU Sits Half Idle While It Generates

AI Is Running Out of Text to Train On. Annotation Is the Fix.

‌

Join the newsletter

Thinking out loud about
Agentic AI

Latest Article

Agent Evals: Your Agent Passed the Test and You Still Don't Know If It Worked

All Articles

Reviewing AI-Generated Code Is a Different Job

Reward Hacking: Why Coding Agents Pass Tests They Should Fail

Tool Poisoning: The Supply-Chain Attack Your AI Agent Can't See

Deep Agents: Why Your Agent Needs a Plan, a Filesystem, and Sub-Agents

Fine-Tuning Your First LLM with Unsloth (Yes, on a Free GPU)

OpenClaw: I Put a Self-Hosted AI Agent in My WhatsApp. Here's the Catch.

n8n for Developers: When to Stop Writing Glue Code

Stop Calling OpenAI Directly: Build an LLM Gateway Instead

LLM System Design: Why Your Backend Instincts Get You Halfway

Stop Eyeballing Your Prompts: Evaluation-Driven Development for LLMs

LLM Inference: Why Your GPU Sits Half Idle While It Generates

AI Is Running Out of Text to Train On. Annotation Is the Fix.

Thinking out loud aboutAgentic AI

Latest Article

Agent Evals: Your Agent Passed the Test and You Still Don't Know If It Worked

All Articles

Reviewing AI-Generated Code Is a Different Job

Reward Hacking: Why Coding Agents Pass Tests They Should Fail

Tool Poisoning: The Supply-Chain Attack Your AI Agent Can't See

Deep Agents: Why Your Agent Needs a Plan, a Filesystem, and Sub-Agents

Fine-Tuning Your First LLM with Unsloth (Yes, on a Free GPU)

OpenClaw: I Put a Self-Hosted AI Agent in My WhatsApp. Here's the Catch.

n8n for Developers: When to Stop Writing Glue Code

Stop Calling OpenAI Directly: Build an LLM Gateway Instead

LLM System Design: Why Your Backend Instincts Get You Halfway

Stop Eyeballing Your Prompts: Evaluation-Driven Development for LLMs

LLM Inference: Why Your GPU Sits Half Idle While It Generates

AI Is Running Out of Text to Train On. Annotation Is the Fix.

‌

Join the newsletter

Thinking out loud aboutAgentic AI

Latest Article

Agent Evals: Your Agent Passed the Test and You Still Don't Know If It Worked

All Articles

Reviewing AI-Generated Code Is a Different Job

Reward Hacking: Why Coding Agents Pass Tests They Should Fail

Tool Poisoning: The Supply-Chain Attack Your AI Agent Can't See

Deep Agents: Why Your Agent Needs a Plan, a Filesystem, and Sub-Agents

Fine-Tuning Your First LLM with Unsloth (Yes, on a Free GPU)

OpenClaw: I Put a Self-Hosted AI Agent in My WhatsApp. Here's the Catch.

n8n for Developers: When to Stop Writing Glue Code

Stop Calling OpenAI Directly: Build an LLM Gateway Instead

LLM System Design: Why Your Backend Instincts Get You Halfway

Stop Eyeballing Your Prompts: Evaluation-Driven Development for LLMs

LLM Inference: Why Your GPU Sits Half Idle While It Generates

AI Is Running Out of Text to Train On. Annotation Is the Fix.

Thinking out loud about
Agentic AI

Thinking out loud about
Agentic AI