Agent Engineering

The practical craft of building agents that survive contact with real work: tool design, context, evaluation, failure modes and the boundaries worth enforcing.

A platform crew maintains a shared industrial service spine beneath several independent product work bays
Artificial IntelligenceArchitectureAgent Engineeringleadership
The AI Platform Team Is an Enablement Layer
How AI platform teams can provide reusable harnesses, evaluation systems, source governance, capability controls, and operational patterns without owning every product decision.
Published on
An autonomous excavation vehicle stops where the verified tunnel ends at an unstable unlit void
Artificial IntelligenceAgent Engineeringagent-securityreliability
When Agents Should Refuse to Continue
A practical guide to refusal as an operational state for AI agents when evidence, authority, safety, budget, or task coherence breaks down.
Published on
An autonomous work cell passes a sealed evidence canister through a transfer lock to a human operator
Artificial Intelligencehuman-in-the-loopAgent EngineeringArchitecture
Human Handoffs Are Part of the System
How AI workflows should treat human review, escalation, approval, correction, and ownership as designed transitions rather than interruptions.
Published on
A paper instruction enters a reader while physical valves and interlocks govern access to an autonomous machine behind glass
Artificial Intelligenceagent-securityArchitectureAgent Engineering
The Prompt Is Not the Policy
Why AI policy must live in enforceable rules, source governance, tools, approvals, and evaluators rather than being reduced to instruction text.
Published on
An autonomous aircraft crosses a test course built with crosswinds, ice, blocked passages, and emergency catch nets
Artificial IntelligenceAgent EngineeringreliabilityArchitecture
Design AI Workflows Around Failure Modes
Why reliable AI products should begin with expected failure classes, containment paths, and recovery states instead of only optimizing the happy path.
Published on
Expedition samples and artifacts are arranged on an inspection table and connected to their compartments in an autonomous rover
Artificial IntelligenceAgent Engineeringobservabilityreliability
Evidence Manifests for Agent Work
How agents can make completion auditable by recording required sources, artifacts, checks, decisions, exclusions, and unresolved uncertainty in one durable manifest.
Published on
An operator pilots a submersible through an ocean trench beside a glass-enclosed simulator running a transparent twin
Artificial IntelligenceevaluationAgent Engineeringreliability
Shadow Mode for AI Releases
Why AI features should run beside existing workflows before taking authority, and how shadow mode exposes behavior, cost, evidence, and failure classes safely.
Published on
A mechanical hand uses one small brass token to unlock a narrow door while larger vaults remain sealed
Artificial Intelligenceagent-securityAgent EngineeringArchitecture
Capability Tokens for AI Agents
A practical model for granting AI agents narrow, temporary, auditable authority to use tools, touch data, and perform consequential actions.
Published on
A mechanical governor meters crowded cargo capsules into a single tunnel inside a steam-filled switching hall
Artificial IntelligenceAgent EngineeringOrchestrationreliability
Agent Queues Need Backpressure
How AI work queues should control intake, priority, concurrency, retries, and human review so autonomous systems do not turn demand into uncontrolled motion.
Published on
12345
Page 1 of 5
HomeAboutArticlesProjectsChat