14 posts · page 1/2
Stateless agents fit most tasks. State is the most expensive capability you can add — it doubles your operational surface, breaks your debugging, and rewards exactly the use cases that can't survive without it. Memory, environment control, self-learning. Part 3 of three.
When the cheap tiers run out, the agent has to reach beyond the model itself — into knowledge it doesn't have, tools it can't natively use, or its own previous answer. RAG, tool use, and self-critique: three patterns, three failure modes worth pricing in. Part 2 of three.
Most agent stacks are built one tier too capable for the job. Three of the cheapest architectures — a fixed pipeline, an LLM with rule constraints, and a reasoning loop — solve more problems than the architecture diagrams admit. Part 1 of three.
Every mature framework is a museum of decisions you can't take back. MLX is interesting mostly because it started after the decisions that matter for Apple Silicon were already mistakes — and the things it threw away are the things that were quietly costing the rest of us the most.
Skills aren't a freeform format. The useful ones fit one of five shapes — sequential workflow, multi-MCP coordination, iterative refinement, context-aware selection, domain-specific intelligence. Picking the right shape is most of the design work. Picking the wrong one is most of the bugs.
Most teams ship one of these and call the job done. MCP gives the agent tools. Skills tell it which to use, in what order, with which fallbacks. Without skills, your MCP integration ends with users asking 'okay, what now?'
Skills load in three levels — frontmatter, body, linked references. Most authors compress all three into one. The result is a 12,000-word SKILL.md the agent reads top to bottom whether it needs to or not, paying for the whole thing in context every time.
Every skill lives or dies by a single line in the frontmatter. If the description doesn't trigger, the instructions never load. Most authors spend an hour on a beautiful body and ten seconds on the line that decides whether anyone ever reads it.
Memory, skills, hooks, subagents, plugins. Five layers in a strict order, each earning a specific payoff, each a mistake if you reach for it before a symptom asks for it. The honest part of the ladder is the rungs you don't climb yet.
Most Claude Code shortcut listicles rank commands alphabetically and call it a day. Half of them never earn the keystroke. Eight commands and four keyboard moves cover most sessions — the honest part of the list is which twenty you can skip.
Most AI-built apps die in the gap between 'works on my laptop' and 'ten real users on Monday.' The gap is nine specific lines, almost none of them about the model. Where each line sits, and the cheapest move that crosses it.
Implementation got fast, and the cheap alignment touchpoints went with it. Now every 'should we be building this?' fight lands on the PR review — which was never built to carry it. Patterns that shift alignment back to the left without waiting for new tools.