Agents and sub-agents

An agent is a model that calls tools in a loop: act, read the result, decide the next action, repeat until the job is done.

Using an agent is like giving a destination instead of turn-by-turn directions. Instead of "search for the equipment schedule," then "read it," then "find the matching one-line on E-501," you ask which panel feeds the Data Hall 2 racks and the agent works out the steps.

A drawing review as an agent​

Ask the Nomic agent for a coordination review. It does not start marking up sheet one right away. Instead it works through these steps:

1. Survey. Reads the sheet index, works out disciplines and sheet counts, notes what is missing.
2. Scope. Reports what it found, then asks which review type you want (Light, Thorough, or Comprehensive) and lists the issue classes and disciplines it would cover. Your answers set the scope.
3. Review. Dispatches a sub-agent for each issue class, meaning a category of problem such as door and hardware conflicts or grid misalignment. Each sub-agent searches, reads, renders, and records findings with markups. Each finding is checked by the drawing judge before it is recorded.
4. Results. A findings table and an annotated PDF.

Each step follows from the result of the one before, and the agent decides what to do next.

Sub-agents​

Some tasks are too large for one context window, and others split naturally into separate parts. An agent can spawn sub-agents: separate agents with their own context and a narrow brief, which report only their findings back to the parent agent.

In a coordination review, one sub-agent owns mechanical, electrical, and plumbing equipment against the schedules, another sub-agent owns door and hardware coordination, and a third owns grid alignment between structural and architectural. Each sub-agent reads only what it needs, which keeps its own context window small. The parent agent assembles the findings table. The Nomic agent also uses an Explore sub-agent to find out what is in the set before the expensive work begins.

Sub-agents cost tokens, often more than the parent agent, because each reads its own slice of the set. A Comprehensive review is expensive for that reason, and it covers the set in more depth for the same reason.

What changes for you​

You set the objective, agree on the scope, and review the output. Once the scope is set, the agent does the searching, reading, and comparing.

Working this way makes several jobs at once possible. A coordination review runs on this week's set while you answer an RFI in another session and a workflow drafts a submittal log in the background. The limit is no longer how many sheets one person can read in a week.

What agents are and are not good at​

Agents are good at tasks with a clear objective and an established pattern: check this set against that standard, cross-reference this submittal against the spec section, list every reference to a product across the spec book, compare two revisions and list what changed, draft the first pass of an RFI response from the documents.

Agents are weak where judgment depends on things not in the documents: design intent agreed on a call, a client's unwritten preference, a site condition nobody photographed. Agents can loop on a failing approach without noticing. Without constraints, they flag every minor inconsistency, even when you wanted only what a contractor would raise as an RFI.

Treat an agent as a well-read junior reviewer who needs a clear brief, forgets between sessions, and works best with checkpoints. Sign-off stays with you.

Delegating​

Start small and well-defined: one discipline, ten sheets, one standard. Learn how it scopes and what it finds. Then widen to the full set, more issue classes, or a milestone review at Extra High effort that runs unattended, and always finish with a verification step of your own.

Module 2​

You now have the foundations: how models work, where they fail, what they cost, what they can see, how they act, and how they split up work. Module 2 is the practice in Nomic: briefing the Nomic agent, setting up a project, scoping a review, checking submittals and answering RFIs, verifying results, and turning a prompt that works into a workflow the whole team can run on every project.

Next module: Working with the Nomic agent