Command Line

Deep Dive

A reference architecture for enterprise agentic AI, in seven layers

Most agent projects fail at the seams between layers, not inside the model. Here is the stack we draw on the whiteboard before writing any code.

2026.10.30

Deep Dive

The agent control plane: what sits between the model and the business

A model can propose actions. Something else has to decide whether they happen. That something is the control plane.

2026.10.28

Deep Dive

Tool access is an authorization problem, not a prompt problem

Giving an agent a tool is giving it a credential. Treat it with the same rigor as a service account.

2026.10.26

Pattern

Durable execution for long-running agents: checkpoints, resumes, and idempotency

An agent task that takes forty minutes will meet a deploy, a timeout, and a rate limit. Plan for all three.

2026.10.24

Pattern

Multi-agent topologies: when a supervisor earns its latency

More agents is not more intelligence. Choose the smallest topology that keeps responsibilities legible.

2026.10.22

Deep Dive

Evals are the contract: writing acceptance tests for non-deterministic systems

If you cannot say what good looks like in a test, you cannot ship the agent, or safely change it.

2026.10.20

Field Note

Agent memory is a data governance surface

The moment an agent remembers something, you have created a new datastore with retention, access, and privacy obligations.

2026.10.18

Pattern

Model routing and the cost envelope: budgets per run, not per token

Finance does not care about your token price. They care that a runaway loop cost what a mid-sized project does.

2026.10.16

Field Note

Human-in-the-loop is an interface contract, not a fallback

Approval screens that nobody can evaluate are theatre. Design review as a first-class part of the system.

2026.10.14

Pattern

Agent observability: trace the decision, not just the call

An HTTP trace tells you a model call took two seconds. It does not tell you why the agent refunded the wrong customer.

2026.10.12