Why LLMs Fail: The 5 Root Causes Behind Every Agent Failure Mode
LLM agents fail in 22 named ways, but those trace to just 5 root mechanisms in how the model works. Guard the root, cover the whole group.
Read the spec →The Post-Office Pattern: How Agents Should Send Each Other Mail
A centralized mail-routing pattern for multi-agent systems: agents drop messages in one shared outbox, a shell script routes them to inboxes. Scripts for mechanics, agents for judgment.
Read the spec →The Roster: 101 Agents, 11 Teams, One Conductor
The actual scale of a production multi-agent system: 101 agents across 11 available teams, roughly 20 disciplines, and one conductor above all of them.
Read the spec →The Handoff Graph: Why This Agent System Isn't a Pipeline
Eleven agent teams declare only what they consume and what they produce. The conductor infers a run order from those declarations, live, every time — a graph that grows by wiring one team, not rewriting a chain.
Read the spec →The Depth Gauge: How Deep a Real Agent Team Actually Goes
Eleven agent teams, sounded three layers deep: the contract the conductor gets, the real headcounts and skills behind it, and how the pieces move.
Read the spec →The Conductor Pattern: How a Real Multi-Agent System Actually Coordinates
A growing roster of independent agent teams, one conductor that never reads their internals, and the rule that makes it work: trust is built by refusing to look.
Read the spec →The Anatomy of an Agent Prompt: 5 Parts Most Prompts Are Missing at Least Three Of
Identity, boundaries, procedure, voice, and self-report tell a model how to behave under pressure. Most agent prompts only have two of the five.
Read the spec →Field Report: What a Production Multi-Agent System Actually Taught Us
A real multi-agent system's failure, what it revealed, and why the specifics stay redacted. Proof of operating depth, not a blueprint to copy.
Read the spec →Stop Listening to the Edges: AI Takes Are Noise, the Middle Is a Conversation
One edge says AI does the thinking for you. The other says AI does no thinking for you. Both absolutes, both wrong — the signal is the conversation in the middle.
Read the spec →Agent Architecture: The Org Chart Every Multi-Agent System Already Has
Every multi-agent system is secretly hub-spoke, mesh, or hierarchy. Each shape has a name and a predictable way it breaks. Pick on purpose.
Read the spec →The Memory Ladder: Files, Database, RAG, or Knowledge Graph?
When to use files, a database, RAG, or a knowledge graph for AI and agent memory. Four rungs, and a climb trigger for each.
Read the spec →22 LLM Agent Failure Modes (and the Prompts That Guard Against Them)
22 ways LLM agents fail, each with the drop-in prompt that counters it. Guardrails for agentic systems.
Read the spec →ActiveAI — ActionMailer for LLM Calls in Rails
ActiveAI does for LLM calls what ActionMailer does for email — a Rails gem with generators, agentic loops, and full ActiveSupport instrumentation built in.
Read the spec →