9 part series

An 8-part playbook that takes Cole Medin's agentic AI teaching and applies it to IBM Maximo teams: what an agent really is, prompts and tools, RAG over Maximo data, MCP tool design, security and guardrails, the AI Layer for agentic coding, local and self-hosted models, and running agents in production, in the field and overnight.

A plain-language definition of an AI agent for IBM Maximo teams: the four core components, the reason-act-observe loop, how an agent differs from a Maximo workflow, an escalation and an automation script, which Maximo jobs are agent-shaped and which must stay deterministic, where the MAS 9.2 Maximo Assistant and MCP server fit, and a decision table for choosing between agent, workflow and autoscript.

A runtime agent that works against IBM Maximo succeeds or fails on three things it reads on every turn: the system prompt, the tool descriptions, and the data its tools return. This playbook applies Cole Medin's prompting, tool-design and context-management lessons to mxapiwo, mxapiasset and mxapiinventory: positive instructions, dynamic prompts that inject site and security context, oslc.select diets against context rot, structured triage outputs, and prompts versioned in git.

Not every Maximo question is a RAG question. A practitioner's guide to choosing retrieval strategies for Maximo data: query structured records live, embed narrative text, use keyword and hybrid search for part numbers and asset tags, hierarchical RAG for OEM manuals and doclinks, graphs for asset, location and failure hierarchies, metadata filters for site security, and delete-before-insert when a work order changes.

A practitioner's playbook for exposing IBM Maximo to AI agents through the Model Context Protocol: what MCP really is, the too-many-tools problem, the MAS 9.2 native MCP server, TMG's Max_mcp and Max_Interfaces, how to design a Maximo tool (names, descriptions, input schemas, oslc.where safety, pagination, idempotent writes), stdio versus streamable HTTP, API keys versus OAuth, and when skills or code execution beat MCP.

A defensive playbook for AI agents on IBM Maximo: map the lethal trifecta onto work orders, service requests and vendor email, walk through a prompt injection that arrives in a service request, and build layered guardrails with least-privilege security groups, one API key per agent, read-only defaults, human approval for status changes, egress control, eAudit evidence and data residency choices.

How a Maximo development team builds an AI Layer on top of its customisation repo so any coding agent (Claude Code, IBM Bob, Codex) writes autoscripts, Java MBO extensions, DBC scripts, MAF apps and integrations the house way: lean global rules, per-artifact rules, a command library (/prime, /plan-autoscript, /validate), hooks, validation against a dev Manage instance, a separate reviewer, autonomy levels per artifact, parallel work, and system evolution in a 15-year-old customised codebase.

How utilities, nuclear operators, defence and public-sector Maximo teams decide what to run on their own hardware: privacy tiers mapped to Maximo data, a per-component local-vs-cloud decision tree, GPU sizing next to MAS on OpenShift, the context window trap with long work order histories, a cost crossover table, and multi-model routing with a small local classifier and a larger reasoning model.

The last mile of the Agentic AI Playbook for Maximo: packaging an agent as a container and API endpoint next to MAS, per-user API keys and multi-tenancy, observability and cost per run, a hands-free voice agent for technicians reading job plans and safety plans, heartbeat and event-driven agents fed by publish channels and Kafka, long-running loops with a human gate, and a production readiness checklist.