Advanced
Architecture, evaluation, governance. The pieces operators come back to once an agent system has been running long enough to surface hard questions.
These pieces are for operators running long-lived agents at scale. They cover architecture decisions, evaluation under uncertainty, governance, and the open questions.
Concept Briefs
Short explainers on loops, graphs, memory, prompting, and the core vocabulary of agentic systems.
-
Agents and Robotics: Physical Workflows
A software agent that sends a wrong email is reversible. A robot that drives into a wall is not. This piece covers what changes when the agent's actions touch the physical world — the failure costs, the latency budgets, the safety patterns, and what current AI can and cannot do.
CONCEPTS -
Evaluation, Safety, and Governance for AI Agents
Agents fail in ways classical software does not. Eval has to catch open-ended, multi-step, tool-mediated behavior. Safety has to bound runaway loops, prompt injection, and tool misuse. Governance has to make the system auditable. This piece is the advanced operator's checklist for all three.
CONCEPTS -
Future of Agent Architectures
Today's agent architecture is mostly language model plus memory plus tools plus loop. It works and will not be the final form. This piece is a forward-looking map of where the architecture is heading — what is being researched, what is being deployed, what is speculative, and what remains unknown.
CONCEPTS -
Sandbox and Terminal Safety
An agent that can run shell commands is an agent that can destroy the file system, leak secrets, or call external APIs in ways the operator did not authorize. The sandbox is the discipline of bounding what the agent can do, what it can read, and what it can write. This piece is an operator's guide to the threat model, the five defensive practices, and the recovery playbook when the sandbox fails.
CONCEPTS -
Agentic RAG: When Retrieval Thinks Before It Answers
Traditional RAG treats search as a static lookup. Agentic RAG lets the agent plan what to fetch, choose between retrieval tools, and re-query until the answer holds up. Here is what changes when retrieval itself becomes a loop, and when it is worth the added complexity.
CONCEPTS -
Multi-Agent Orchestration: When One Agent Becomes a Team
The next failure mode after a working single-agent loop is the agent that tries to do everything at once. Multi-agent orchestration splits one loop into a coordinator and specialists with explicit handoffs and a shared contract. Here is when it makes sense and how the three patterns work in practice.
CONCEPTS
Workflow Reviews
Pattern reviews: what a given workflow is good for, where it breaks, and how to improve it.
-
Effective Cron Design: Idempotency, Pacing, Retries, and Condition Triggers
An effective cron is one you can run twice without breaking anything. The disciplines are idempotency, pacing, retries, condition triggers, and observability. Each one is cheap to add; each one is expensive to skip.
WORKFLOWS -
Graph-Driven Pipelines: Blender, Game Engines, and Printers
When Blender hands off to a game engine, a 3D printer, or another tool, the handoff has shape. State graphs are how you make that shape explicit, debuggable, and recoverable.
WORKFLOWS -
Long-Context Agents: Designing Workflows That Survive a Million Tokens
Long context changes what an agent can hold in mind, but it does not remove the need for good memory, good chunking, or good governance. This is the working set of patterns for designing workflows that use long context well — and the failure modes that arrive when long context is used as a substitute for design.
WORKFLOWS -
Designing Safe Render Loops for Blender
A render queue should be a loop that respects budgets, fails honestly, and never modifies the source scene. The discipline is in the bound — not in the cleverness of the agent.
WORKFLOWS -
State Graphs in Practice: Beyond Loops and Pipelines
Loops are how an agent makes progress. State graphs are how an operator builds a workflow that survives being long, branched, and shared with other agents. This article is about the move from loop-shaped systems to graph-shaped systems — what the new shape buys you, what it costs, and how to keep the graph from collapsing into spaghetti.
WORKFLOWS