Andreas Nissen

AI Agents: The Definitive Guide

Mon, 01 Jan 0001 00:00:00 +0000

The original companion repository for AI Agents: The Definitive Guide.

Runbook Relay

Mon, 01 Jan 0001 00:00:00 +0000

The problem

Incident dashboards are dense, stateful, and consequential. Screenshot-driven automation has to infer what a chart means, which change is selected, and whether an operator approved execution. That is too much ambiguity for an operational control surface.

What I built

Runbook Relay exposes narrow WebMCP tools for reading an incident, comparing mitigations, staging a change, executing an approved change, and resetting the simulation. The tools and the human interface use the same state transitions.

Mistral Reliability Lab

Mon, 01 Jan 0001 00:00:00 +0000

The problem

A model API quickstart proves that a request can return text. It does not prove that tool execution is bounded, failures are handled deliberately, model behavior is testable, or dependency risk is visible.

What I built

Mistral Reliability Lab is a modular Python reference with a reviewer-first Streamlit interface, a FastAPI surface, opt-in OpenTelemetry tracing, and support for both Mistral’s API and local Ollama models. Its credential-free preview lets a reviewer inspect the interaction model without supplying an API key.

Context and Control Are Different Layers

Sun, 30 Aug 2026 00:00:00 +0000

Enterprise AI agents need context and control.

Context without control is unsafe. Control without context is blind.

They belong in the same architecture, and they should not collapse into the same layer. A missing relationship degrades an answer. A missing authorization check can trigger an action the user never permitted. Those failures have different consequences, owners, and tests.

Context answers what the agent should know

Context is the evidence available for reasoning. It can include documents, structured records, tool results, conversation state, identity attributes, and the relationships between them.

From Screenshots to Governed Tools

Sat, 29 Aug 2026 00:00:00 +0000

An agent can use a website by looking at pixels, clicking controls, and reading the result. That is powerful because it works without changing the application.

It is also the least explicit way to automate a consequential interface.

The agent has to infer what a chart means, which control maps to which action, whether the page changed, and whether an approval actually exists. The human sees a visual interface. The agent reconstructs a hidden API from appearance and behavior.

An Ontology Is More Than Context

Fri, 28 Aug 2026 00:00:00 +0000

Context is what an AI system can see for the current task.

An ontology defines what that information means.

The terms are related, and they are not interchangeable. A long prompt can contain plenty of context without a stable model of the domain. An ontology can describe the domain even when only a small part of it is retrieved for one request.

Context is task-specific evidence

Context is assembled for a particular interaction. It can include instructions, documents, records, conversation state, retrieved passages, tool results, and user preferences.

Free AWS Builder Center Sandboxes Make Hands-On Learning Easier

Thu, 09 Jul 2026 00:00:00 +0000

Want to go hands-on with Amazon Bedrock, agentic AI workflows, or AWS Transform custom without first setting up an AWS account and payment method?

AWS Builder Center now offers workshops with free, pre-provisioned sandbox environments. You can request an eligible environment, work through the lab, and let the temporary account expire when the session ends. The current sandbox window is eight hours, which is enough for a focused workshop without turning account administration into a separate project.

The Hard Part of Agentic AI Starts After the Demo

Mon, 15 Jun 2026 00:00:00 +0000

The hard part of agentic AI is not building the agent.

It is everything that happens after the demo works.

A prototype can take a few days. Production is a different discipline. Agents reason in loops, invoke tools, change external state, keep memory, and sometimes produce different answers to the same request. The request-response playbook that teams have used for conventional applications does not cover all of those behaviors.

The AWS Well-Architected Agentic AI Lens gives teams a framework for closing that gap. It adapts all six Well-Architected pillars to the way agentic systems actually behave: operational excellence, security, reliability, performance efficiency, cost optimization, and sustainability.