AI Agents for PCB Specification and Review

Specifying a board consists mostly of work other than drawing. It consists of selecting parts that are genuinely in stock, reading the datasheets closely enough to establish what each component requires on either side of it, and confirming that the whole assembly is coherent before a production run is paid for.

The Shape of the Work

Observed directly, the time divides into three activities. The first is search: identifying candidate parts that meet the requirement, arrive in a package the assembler can place, and are purchasable in the required quantity at a price the product can carry. The second is reading: a modern power management integrated circuit ships fifty pages of datasheet, and the twelve sentences that determine whether a design functions are distributed across all fifty. The third is checking: a regulator enable pin requiring a pull-up to a rail that comes up before it, an ADC reference that cannot exceed its supply, a level shifter oriented incorrectly for the bus it serves.

None of that is unskilled. All of it is bounded, checkable, and highly repetitive, which describes the class of work where a language model contributes value and does not describe the class where a language model should be trusted.

System Overview

Conductor is a Sunlab Digital product. A lead engineer agent takes the intake, an orchestrator plans the work, a researcher performs the reading, and a system architect assembles the bill of materials and validates it against live JLCPCB and LCSC stock and against the datasheets themselves. The system will also review an existing schematic, where every finding must clear two gates before it is displayed: an observation in the design under review, and a corresponding requirement in the manufacturer’s own document.

The division of labor between those agents matters less than the constraint each operates under. Every agent has a narrow task with a defined output shape, so an answer can be rejected structurally rather than argued with. The researcher does not return prose about a component. It returns claims, each carrying the document, the page, and the passage it derives from. A claim without a citation is discarded rather than downgraded.

Separation of Reasoning and Effect

The governing rule is that agents reason and emit intent, and deterministic code validates, persists, applies, and audits every side effect. Nothing a model produces reaches a schematic without passing through code that can be tested.

In practice the model never writes a netlist. It emits an instruction: connect this pin to that rail, add this decoupling capacitor at this location, set this resistor to this value. A validator then determines whether the instruction is legal. Voltage domains must agree. A pin declared as an output cannot be tied to another output. A part absent from the resolved, stock-checked bill of materials cannot appear in a connection. Where an instruction fails, the failure returns to the agent as a specific mechanical complaint rather than as an unqualified retry.

The property this arrangement produces is that system correctness does not depend on the model being right. It depends on the validators being right, and validators are ordinary software: readable, arguable, and testable rule by rule. Where the model improves, the output improves. Where the model produces an unsupported claim, the output is rejected. Both behaviors are required, and a system that permits the model to write directly exhibits neither.

Datasheet Ingestion and Retrieval

Datasheets are among the most difficult documents in engineering to parse and among the most important to parse correctly. They are multi-column PDFs in which the meaning resides in tables spanning pages, in footnotes attached to those tables, and in absolute maximum ratings printed alongside recommended operating conditions and carrying an entirely different meaning.

Naive chunking destroys that structure. A table row separated from its header row and its unit column is worse than absent data, because it presents as data. Ingestion therefore preserves table structure and binds every extracted value to its provenance: which part, which document revision, which page, which cell. Revision matters more than is generally assumed. A design validated against Rev C of a datasheet is not validated against Rev E, and the errata between them are precisely the class of issue this tool exists to identify.

Schematic Review

Review is the mode with the highest cost of error. A review tool producing forty findings of which thirty are noise does not return time to an engineer. It adds a triage task to the week and trains the engineer to stop reading the output.

The two gates follow from that. A finding requires an observation in the design under review and a quoted requirement from the manufacturer’s own document. A convention carrying no citation is suppressed. A datasheet requirement the design already satisfies is suppressed. What remains is a shorter list than comparable tools produce, which is the intent.

Defined Scope

Conductor does not perform layout, and extension into layout is not planned. Placement and routing are where the physics ceases to be a set of contracts between pins and becomes a continuous problem with electromagnetic consequences, and the accurate statement of the current state of the art is that an engineer with a capable tool performs it better.

The system also does not sign anything off. The output is a bill of materials, a schematic, and a written record of why each part is present and which document justifies it, delivered to an engineer who remains accountable for the board. The value is that the review begins with the reading already complete, not that the review is skipped.

Architecture for Agent Systems

The first decision in an agent system is which half of it is permitted to be wrong. Here it is the reasoning half, deliberately, and every structural choice follows from that: the narrow output shapes, the citation requirement, the validator between intent and effect, and the audit log that allows a component selection to be reconstructed months later. Systems built without that decision tend to demonstrate well and prove undeliverable across a quarter.