Agent Architecture Map

Writes one complete Agent Role Definition for a single proposed node, using Paul Cheek's eight-section worksheet. Produces goal, never list, guardrails as numbers, named escalation, memory, metrics, and action log, plus the justification for the node existing at all. Defines one node per invocation so each gets full attention.

Adversarial review of a finished architecture before it is rendered. Ranks the six layers strictly one to six, audits every action the agents can take against their human relationship, verifies the never lists are enforced rather than wished, and rules on whether this should be one agent or a team. Assumes the architecture is wrong and tries to show it.

Deep research on a named organization before its architecture is designed. Establishes what it does, how it makes money, its systems of record, its regulatory constraints, its recent AI posture, and where the named use case would actually sit. Every finding carries a source. Never invents facts to fill a gap.

Places the people and draws the reporting lines. Decides which humans appear as carbon nodes, assigns RACI between every person and every agent, names who owns each agent's expenses and who owns its inference cost, sets approval gates, and routes escalations to named individuals. Enforces one accountable person per agent and a named budget owner for both cost lines.

Designs the memory layer. Decides what persists between runs, at what scope, for how long, and who may read or write it. Distinguishes working state from institutional memory, and keeps shared reads cheap while making shared writes a deliberate decision.

Decomposes a business process or workflow into steps, decision points, handoffs, waiting time, and the systems each step touches. Produces the work breakdown every other agent builds against, and the as-is all-human baseline the finished architecture is measured against. Does not design agents or assign tools.

Owns the grants. Builds the tool, memory, and budget inventory, assigns every agent-to-resource edge a level on the read/draft/write/commit ladder (owns/draws for budgets), writes the justification and the concrete actions for each, and cuts every edge nobody can defend. Enforces least privilege as a deletion exercise, not a labeling one.

Maps the possibility space for a use case rather than jumping to one answer. Produces several complete, materially different architectures for the same outcome, each with what it buys and what it gives up, so the choice is made deliberately instead of by default.

Design a multi-agent architecture for any business process and render it as an interactive HTML page on the No One Works Here design system. Draws the whole network of human reporting nodes, orchestrators, agent teams, sub-agents, and parallel fan-outs, with the tools each agent may touch and the memory stores it may read or write, every permission least-privilege and every node and edge carrying a written justifica…