Agentic AI Knowledge Hub
How AI workers are designed, controlled and measured — and how to tell which part of your work should use one.
Read the five steps below in order if you own the work. If you have to be satisfied this is safe before it touches production, the technical deep dives underneath carry the architecture.
Start here
For the people who own the work
Five steps, in operational language, ending with how the result gets measured. No architecture vocabulary required.
- The problem
Why work falls through
Where queues, handoffs and waiting build up, and why adding people has stopped fixing it.
Foundation · 8-minute read
Coming soon
- Business method
The Nimbus180 AI Transformation Compass
Start with the business outcome and redesign the work before choosing any technology.
Foundation · 15-minute read
Coming soon
- Choosing the mechanism
When to use a worker, a workflow or a rule
Judgment belongs to agents, control belongs to workflows, and most rules should stay deterministic.
Core · 12-minute read
Coming soon
- Division of labor
How humans and AI workers divide the work
What the worker carries, what stays with your people, and where the handoff happens.
Core · 10-minute read
Coming soon
- Control
Privacy, control and measurable outcomes
What the worker may see and keep, what gets recorded, and how the result is measured.
Core · 12-minute read
Coming soon
Technical deep dives
For architects, engineers and security reviewers
The engineering underneath: context, memory and retrieval, tool permissions, the runtime loop, and how far autonomy should go. Read in order, or go straight to the one you are evaluating.
- Agent basics
Anatomy of an AI Agent
An agent is a controlled system around a model, not the model by itself.
Foundation · 12-minute read
Coming soon
- Input
Context & Prompts
Design what the model can see, and how instructions are prioritized against untrusted data.
Core · 12-minute read
Coming soon
- Knowledge
Agent Memory & RAG
Separate conversation state, durable memory, retrieved evidence and the system of record.
Core · 14-minute read
Coming soon
- Action
Tools, APIs & MCP
Give the agent safe, discoverable, permissioned ways to act on external systems.
Core · 14-minute read
Coming soon
- Runtime
AI Agent Lifecycle
Bound the runtime loop with state, checkpoints, retries, evaluation and escalation.
Core · 14-minute read
Coming soon
- Maturity
The Five Levels of Agentic AI
Scale autonomy and governance together, using the lowest level that delivers the value.
Core · 12-minute read
Coming soon
- Advanced / optional
Reinforcement Learning for AI Agents
Consider learned policies only when simpler approaches are exhausted and feedback is measurable.
Advanced · Optional · 14-minute read
Coming soon
Required at every step
Security, evaluation, observability, human oversight and privacy are engineering disciplines that run through all nine steps, not modules added at the end.
- Security
- Identity, least privilege, tool allowlists, prompt-injection resilience and audit records around every action the agent can take.
- Evaluation
- Model, agent, system and business measures, run against a versioned suite before any change to prompts, models, retrieval, tools or policy.
- Observability
- Traces, decisions, tool calls, cost and outcomes — inspectable by a person who is not an engineer, without exposing data they do not need.
- Human oversight
- Inform, review or approve, chosen by consequence rather than added vaguely as a human in the loop.
- Privacy and minimization
- Purpose, classification, retention and deletion across context, memory, retrieval, tools, logs and feedback.
Our position
The goal is not to build the most autonomous agent. The goal is to use the lowest level of autonomy that reliably, safely and economically delivers the required business value.
Discuss your workflow
The fastest way to use any of this is against one workflow you already run. Bring us that workflow and we will map it with you.