Agents & AI Engineering
AI engineering is making a model do the same correct thing twice: orchestration, evaluation, and the failure modes that appear once an agent has permissions. Heavy emphasis on security, because an agent that can act is an agent that can be exploited.
Latest in this section
Tools for this →AI Oversight Without Reviewing Everything
AI oversight that scales: cheap first-line signals, escalation to costly review, a human on every irreversible call. The three tiers and how to build them.
Oct 31, 2026 · 15 min read
Reward Hacking: When AI Passes the Test, Not the Task
Reward hacking is when an AI exploits its success measure, not the task: editing the tests, gaming the grader. What it is, what it isn't, and how to catch it.
Oct 31, 2026 · 11 min read
Why an AI's Explanation Isn't Proof of Its Reasoning
AI explanation faithfulness: a convincing explanation is not proof of its reasoning. What counts as evidence, what Anthropic found, and how to check it.
Oct 31, 2026 · 11 min read
Open Models, Closed APIs, and the Right to Inspect AI
To inspect AI models from the inside you need the weights, which a hosted API never gives you. What open weights change, what they don't, and what to ask for.
Oct 31, 2026 · 10 min read
The AI Inspectability Checklist: 10 Vendor Questions
Ten AI vendor explainability questions with the answer that counts as evidence for each. A buyer's checklist, not a certification. Free one-page PDF.
Oct 31, 2026 · 10 min read
AI Interpretability: The Black Box May Be Opening
AI interpretability is moving: neurons described, pathways traced, behavior changed on purpose. What it proves, what it doesn't, and what you can check now.
Oct 30, 2026 · 15 min read
Neural Networks Are Already Clean: What the Essay Says
Neural networks are already clean, says Goodfire's Nick Cammarata. The claims, the numbers kept apart, the caveats, and what the companion paper must show.
Oct 30, 2026 · 11 min read
The Cat and Car Neuron: Solved, or Still Open?
The cat and car neuron is AI's most famous polysemantic neuron. Goodfire reads it as a V detector. Liv Gorton's tests disagree. Both accounts, side by side.
Oct 30, 2026 · 10 min read
Mechanistic Interpretability, Explained in Plain Terms
Mechanistic interpretability without the jargon: neuron, pole, feature, polysemanticity, superposition, circuit, SAE, probe, steering, attribution graph.
Oct 30, 2026 · 10 min read
Does Interpretability Make AI Safe? The Missing Steps
Is AI interpretability enough for safety? No. Legibility is one input. Still missing: prediction, durable fixes, monitoring, and a human who can stop it.
Oct 30, 2026 · 10 min read
How to Estimate and Budget Your LLM Token Costs
LLM token cost equals calls times average tokens per call times your provider's current input and output rates, priced separately and always changing.
Oct 23, 2026 · 7 min read
How Much Does an LLM API Cost? (Input vs Output, Explained)
LLM API cost splits into two rates: a lower one for the text you send in and a higher one for what the model generates back. Both keep changing.
Oct 23, 2026 · 6 min read
What Is a Token in AI? (And Why It Costs You Money)
What is a token in AI? A word or a piece of one: the unit an LLM reads and writes in, and the unit nearly every provider bills you by, per call.
Oct 23, 2026 · 6 min read
Input vs Output Tokens: Why Your AI Bill Is Higher Than You Expect
Input vs output tokens are billed at different rates, with output priced higher. Why the split exists, and how re-sent chat history inflates the bill.
Oct 23, 2026 · 7 min read
How to Reduce LLM Token Costs (Without Wrecking Quality)
Reduce LLM costs by trimming your system prompt, sending only relevant context, capping chat history, and reserving your priciest model for real need.
Oct 23, 2026 · 6 min read
AI Code Review: What to Automate and What to Keep Human
AI code review handles PR descriptions, security smells and breaking-change checks well. Exploitability, design fit and the ship call still need a person.
Oct 13, 2026 · 6 min read
PR Hygiene: The Checks That Belong in Every Pull Request
PR hygiene blocks four things before review: conflict markers, debug statements, secrets and skipped tests. Four softer signals get a warning instead.
Oct 13, 2026 · 5 min read
How to Gate Bad Commits Before They Reach CI
A pre-push hook ci gate runs one read-only check on the commits you push, then again in CI, so secrets and debug code stop on your machine, not after merge.
Oct 13, 2026 · 5 min read
How to Write a PR Description Reviewers Actually Read
How to write a PR description reviewers read: say what changed, why, and how to verify it, then name anything breaking before the reviewer finds it.
Oct 13, 2026 · 5 min read
Conventional Commits: A Practical Guide
Conventional Commits writes each message as type(scope): description. Only feat and fix are defined by the spec, and ! or a footer marks a breaking change.
Oct 13, 2026 · 5 min read
Multi-Agent Orchestration: A Practical Guide for Dev Teams
Multi-agent orchestration coordinates specialized agents into one flow. Add agents only when a task truly needs splitting, not because more sounds better.
Oct 10, 2026 · 8 min read
Agent Orchestration Patterns: Sequential, Parallel, and Hierarchical
Agent orchestration patterns come in three shapes: sequential, parallel and hierarchical. Pick the simplest one, then put a gate at every hand-off.
Oct 10, 2026 · 7 min read
Single-Agent vs Multi-Agent: When You Actually Need Orchestration
Single agent vs multi agent: start with one agent, which is cheaper and easier to debug, and split only for a concrete reason you have measured.
Oct 10, 2026 · 6 min read
The Orchestrator-Worker Pattern, Explained
The orchestrator worker pattern has one lead agent plan and delegate to specialist workers, and it needs a hard limit on delegation rounds or it loops.
Oct 10, 2026 · 6 min read























