LEAD-01-021-P · Primer — The Enterprise Reference Architecture

The idea in one sentence

An enterprise agent is not one clever prompt but a pipeline of named stages — goal, plan, task queue, tool selection, execution, memory, evaluation, retry, human approval, done — that you can point to on a whiteboard.

Why it exists

Demos hide the machinery: a chat box goes in, magic comes out, and nobody can say where a failure lives. Enterprises need the opposite — an explicit assembly line where every stage has an owner, a log, and a place to fail safely — the same reason a hospital has intake, triage, treatment, and discharge instead of one heroic doctor doing everything at once. When you can name the stages, you can insure, audit, and improve them.

How it works

A goal enters as a clear brief (this is CARE's Clarity). 2) A planner decomposes it into steps (GPS: Goal, Proof, Steps). 3) Steps land in a task queue so work is resumable, not held in the model's head. 4) For each step the agent selects a tool (search, database, API) — the Resources in CARE. 5) An execution engine runs the tool and captures the result. 6) Memory records what happened so later steps and later runs can use it. 7) An evaluator checks the output against the proof standard — the hardest, most-skipped stage. 8) On failure the agent retries or escalates; above a risk threshold a human approves before any real-world action. 9) Only then is the task marked done. Example: an accounts-payable agent moves each invoice through received to extracted to matched to approved to posted — every invoice in exactly one state, every transition logged — so a crash resumes cleanly instead of paying twice.

See it

flowchart TD
  A["Goal + Brief"] --> B["Planner"]
  B --> C["Task Queue"]
  C --> D["Tool Select<br/>+ Execute"]
  D --> E["Memory"]
  E --> F["Evaluate"]
  F -->|fail| G["Retry or<br/>Escalate"]
  G --> C
  F -->|high risk| H["Human Approval"]
  F -->|pass| I["Done"]
  H --> I

In practice

Finance close: a month-end close agent takes the goal 'reconcile intercompany balances,' plans the sub-tasks, queues each entity pair, pulls ledgers via tools, reconciles, flags variances over a threshold to a controller for approval, and posts the rest — with every step time-stamped for the auditors.

The one thing to remember

If you cannot point to the stage where an agent failed, you have a demo, not an architecture.