AI design agents: brief the task and accept the actual file

Look for executed operations

A design agent may inspect a file, select objects, calculate a result, create an output and inspect it again. A chat interface can also expose tools, so the distinction is the work performed and the evidence returned. A long explanation does not establish that a drawing was opened or changed.

Anthropic's engineering description of agents distinguishes model-directed tool use from predefined workflows. For a design team, the practical question is whether that autonomy stays inside an agreed task and produces something reviewable.

Ask what tools are available in the actual session. Ordinary chat, exported data, a script and a configured CAD connector provide different access. A tool's existence in another product does not mean this agent can use it.

Define scope, actions and stop conditions

A task brief should identify input files and revisions, the processing region, units, coordinate reference and expected outputs. State which design decisions are already accepted and which changes need clarification.

Do not ask the agent to repair all gaps merely to make a calculation run. Closing the nearest endpoints can change an intentional opening. Require the location and proposed action when repair affects the result.

For regulatory tasks, identify the actual jurisdiction and source edition. Measurement and source applicability are separate evidence streams. Code-research checks describe the information needed before a threshold enters a finding.

Worked example: a surface-area register

Suppose a site designer supplies a DWG with building outlines, circulation boundaries and approved surface types. Two contours have no confirmed classification. The first deliverable should be an object register and two questions, rather than a completed drawing based on guesses.

Inspect revision C and list the surface contours in the marked region. Preserve all boundaries. Assign types only from the approved table where the mapping is explicit. Keep ambiguous contours outside typed totals and show their IDs and areas separately. Return units, object selections, calculation method and exclusions. Ask for clarification before creating a presentation drawing.

The designer confirms classifications and checks the region and boundaries. If the accepted quantities are 2400 m² of vehicle paving and 800 m² of pedestrian paving, the typed total is 3200 m². These illustrative numbers need an independent check of representative contours and the full accounting boundary.

An equal overall total is insufficient if two contours were assigned to the wrong types. Review the category subtotals and check for overlapping or omitted areas. If the workflow creates a drawing, open the actual output and compare its hatches, marks and schedule with the approved register.

Require evidence at each important step

The input inspection should report missing references and unsupported entities. The calculation should retain selected-object IDs and rules. A created file should come with its name, revision relationship and change list. A self-check should say what was tested and what remains unresolved.

Self-checking is useful but does not become independent merely because it runs twice. Repeat a critical measurement using another method or compare against a known control. Two calculations using the same wrong boundary can agree perfectly.

For Revit-originated work in ARAI, use the required DWG/PDF exports and confirm their contents. Do not substitute a claim of native RVT inspection. For reconstruction from PDF or images, first verify that the installed reconstruction capability is available, then define calibration and geometric acceptance separately from text recognition.

Make corrections specific and review their reach

A useful correction names the object, actual state, expected state and source: “Contour P-12 is pedestrian paving in the output; the approved table assigns vehicle paving.” After correction, check that contour and every affected subtotal.

Ask for the changes relative to the previous output. If the agent regenerated the whole drawing, widen review to match the possible changes. A corrected label should not silently change geometry.

Stop when repeated attempts do not resolve the cause. Return to the input, operation or tool support instead of issuing the same broad instruction again. Define a practical limit based on review effort and the importance of the task.

Measure useful work and try ARAI

Track input preparation, execution, review, corrections and calendar waiting separately. Agent runtime alone does not establish labour savings. Compare accepted outputs from the same kind of task before drawing conclusions about productivity.

Prepare a bounded task in ARAI. Review the unsent draft and attach the current file. Start with one supported drawing operation whose result you can independently check.