AI Engineering

Reference material on how production AI systems are actually built.

Four deep sections covering agent architecture, the harness that runs an agent, the model families available to it, and where bots and assistants fit. Written for the technical and operational leaders who have to own these systems after they ship.

Agents

What an agent actually is in production: a planning loop bounded by tools, state, and stop conditions — and the step-level failure economics that decide whether one is viable.

Agent harnesses

The runtime around the model: tool registry, context assembly, policy enforcement, sandboxing, retries, tracing, and the replay artifact your auditors will ask for.

Models

Reasoning, balanced, fast, long-context, and multimodal families compared on capability, latency, and relative cost — plus how routing between them is decided.

Bots & assistants

Where scripted bots, retrieval assistants, copilots, and autonomous agents sit on the autonomy-versus-oversight matrix, and how to choose deliberately.

Read

Engineering principles

The constraints we hold constant across every engagement.

Determinism where it matters

Models handle language. Control flow, arithmetic, authorization, and writes stay in code that can be tested and replayed.

Every decision is traceable

A run emits the plan, the inputs, the tool calls, the model chosen, and the reason. Regulated clients need to reconstruct decisions months later.

Oversight is a design parameter

Autonomy is set per step, not per system. Mutating and irreversible actions carry explicit approval gates.

Right model per step

Workflows are decomposed so each step runs on the cheapest model tier that clears its quality bar within its latency budget.

Exercise the components

Every concept here has a callable endpoint.

The planner, model router, classifier, and workflow validator described in these pages run as public endpoints you can call from the browser or your terminal.

Open the live consoles