ItInsightsWorld AI Applications Review: Agentic Engineering Tools — CADFit, Celus, ChipAgents, Circuit Mind AVE and Cadence Maestro

By James Hartley · ItInsightsWorld, AI Applications Desk

Applied-AI owners are no longer being sold a smarter answer box. The latest Engineering Paper batch of design and simulation announcements is best read as a deployment map: several vendors are shipping software that drafts a plan, drives the engineering tool itself, and holds only for a human approval before executing. That is the point at which agentic AI starts to move cycle time, rework and reviewer load rather than merely impressing a demo audience.

Two threads run through the set. One is CADFit, an open-source mesh-to-CAD project that converts meshes into editable CAD code. The other is a cluster of agentic launches aimed squarely at electronics and chip design — Celus, ChipAgents, Circuit Mind and Cadence — each attaching AI execution to a stage of a flow that until now was staffed by senior engineers.

What actually changed for application owners

Most AI features in engineering software still sit alongside the workflow: they summarise a datasheet, autocomplete a field, or answer questions about a model. The distinction that matters commercially in this round is whether the output is an artefact that downstream tools can consume, and whether the step being automated can be checked against a source of truth. Where both are true, pilots convert to production quickly. Where either is missing, the result is a review burden that engineers quietly reject.

Case in focus: CADFit and the mesh-to-editable-model gap

The rebuild-from-scan handoff is unglamorous and expensive: a mesh arrives from scanning or a legacy file, and a human reconstructs it as a parametric part. Earlier editions of Engineering Paper have tracked CADFit as a mesh-to-CAD effort in the same family as Backflip AI and Autodesk Fusion’s AutoTimeline. The project is led by Ghadi Nehme, a PhD student at MIT.

For a buying committee the interesting part is provenance rather than capability. CADFit is open-source and can be freely used for non-commercial work; the initial research was funded by Altair, now part of Siemens, and Nehme and Siemens have filed for a patent for the method. Treat “usable at no cost” and “clear of intellectual-property questions” as two separate checklist items — they are not the same decision.

The paper’s own framing is deliberately formal: Nehme describes the method as a “hybrid optimization-based CAD reconstruction framework that recovers complex, editable CAD construction sequences from meshes by incrementally fitting and validating parametric operations using geometric feedback.”

The operational translation is short. “[CADFit] basically starts with the mesh and then extracts some key sketches, similar to how engineers would extract them,” Nehme explained. From there the system converts sketches into solids, consolidates them with unions and cuts, compares the reconstruction against the initial mesh, and reiterates until the gaps close.

The loop CADFit runs, step by step

1. Take the mesh as the starting object — not an empty part file.
2. Extract candidate key sketches, mirroring the primitives a human engineer would pull out.
3. Convert those sketches into solids.
4. Consolidate the solids with union and cut operations.
5. Compare the resulting body with the initial mesh and iterate to fill remaining gaps.
6. Express the outcome in CadQuery, the open-source Python library in which CADFit models are written, so the construction sequence stays editable.

Model selection is handled by a deliberately small component. According to Nehme, a simple machine-learning model identifies which sketches are meaningful; it is not essential, but it speeds the process up by a factor of three. For anyone modelling labour effort, that is a workflow-speed statistic, not an accuracy claim — worth keeping separate in a business case.

The declared scope boundary is the most useful line in the whole announcement. “Currently, the operations that CADFit supports are extrusions, revolutions, fillets, and chamfers, and then also union and cut and all the sketch operations,” Nehme said. “But if a shape is very complex, like loft-based or also sweep-based shapes… the math will struggle.” A revised version aimed at those geometries is in preparation.

Integration is the second constraint. Nehme has also built Studio CAD, his own viewer for “human friendly editing of the shapes,” but there are no plugins for common CAD software yet; he is working on one for Siemens NX. In practice that places CADFit today in the internal-tooling or evaluation lane rather than the managed-production lane.

Nehme is explicit about where the release sits on the maturity curve. “The future iteration of this work will potentially be able to be really robust on a very huge set of shapes,” he said. “But this one is already pushing the boundary by a huge amount… This method is able to really extract the key components of the geometry much better than what’s been seen before.”

The electronics-design cluster: plan, execute, verify

Celus — planning before schematic capture

EDA developer Celus has shipped an agentic workflow for its Celus Design Platform, available in beta. The assistant “helps hardware engineers plan the architecture and component work at the start of an electronics design, before schematic capture, and carries out that plan once the engineer approves it,” according to the company’s press release. The approval gate is the product decision worth copying internally: planning is delegated, execution is gated.

ChipAgents — objectives stay human, closure loops do not

ChipAgents, a semiconductor design software developer, says it is expanding its agentic engineering platform across digital and analog chip design. Its description of the division of labour is the clearest statement of the pattern in this batch: “ChipAgents’ full-stack approach enables engineers to define objectives, constraints and critical judgment calls while AI agents carry more of the execution forward, reasoning over design context, operating engineering tools, evaluating results and determining the next action toward closure,” reads the press release.

Circuit Mind AVE — verification as the entry point

ECAD developer Circuit Mind has launched AVE, short for Agentic Verification Engine. The company calls it an “AI application that checks Printed Circuit Board (PCB) schematics against component datasheets, application notes, and each organization’s own design review checklists, catching errors that slip past even careful manual review before they turn into costly board respins.” Of the four, this is the most legible business case: the output is a defect list, the reference material is fixed, and the avoided cost — a board respin — is already understood by finance.

Cadence Maestro — domain-specific models driving existing tools

Cadence has launched Maestro, a portfolio of domain-specific AI models for agentic chip and system design. “Combining our Maestro portfolio of domain-specialized models with leading frontier and open-weight models enables our AI Super Agents to drive Cadence EDA and SDA tools with exceptional skill and efficiency to transform chip and system design,” said Anirudh Devgan, chief executive of Cadence, in the company’s press release. Value here is tied to an existing tool estate: the agents operate Cadence tools rather than replacing them.

A single agentic theme runs through all four: the AI takes over execution inside a toolchain, while objectives, constraints and sign-off stay with named humans.

Side by side: workflow position, checkpoint and readiness

Announcement	Workflow position	What the AI takes over	Human checkpoint	Stated readiness	Application-layer read
CADFit (mesh-to-CAD)	Geometry reconstruction from scans or legacy meshes	Sketch extraction, solid creation, union/cut consolidation, iteration against the mesh	Reviews and edits the construction sequence	Open-source, free for non-commercial work; CadQuery/Python output; no mainstream CAD plugin yet, Siemens NX connector in progress	Evaluation and internal-tooling lane; poor fit for unattended production today
Celus Design Platform agentic workflow	Start of an electronics design, before schematic capture	Architecture and component planning, then execution of the approved plan	Explicit approval before the plan runs	Beta	Early pilot where planning consumes senior engineer time
ChipAgents agentic engineering platform	Digital and analog chip design	Reasoning over design context, operating engineering tools, evaluating results, choosing the next action	Objectives, constraints and critical judgment calls	Expansion across digital and analog announced	Strongest where closure cycles are already tool-driven and measurable
Circuit Mind AVE (Agentic Verification Engine)	PCB design verification	Cross-checking schematics against datasheets, application notes and internal review checklists	Manual review continues; AVE adds a pre-respin second pass	Launched	Lowest-risk first deployment: verifiable sources, known failure cost
Cadence Maestro	Chip and system design across EDA and SDA	Driving Cadence tools through AI Super Agents built on domain-specific models plus frontier and open-weight models	Task definition and acceptance of tool output	Launched as a portfolio of domain-specific models	Adoption depends on the installed toolchain, not on model curiosity

One triage lens to keep: physical AI versus physics AI

In a guest article for Engineering.com, Luminary CEO Pete Schlampp draws the line between physical AI and physics AI, and explores the potential of the latter as “AI that understands physics.” For application portfolios the distinction is operational rather than rhetorical: one category acts in the physical world, the other reasons about governing behaviour. Which side of that line a use case falls on changes the validation you require before it touches a live design.

Pilot triage checklist

• Does the tool emit an artefact your downstream process already consumes — code, a schematic, a check report, a construction history?
• Is there an independent source of truth to check against (mesh, datasheet, application note, internal checklist)?
• Where exactly is the approval gate, and is it a single named role that can block execution?
• What is the failure mode, and is it caught before hardware, before release, or after both?
• Is the deployment status beta, launched, or research release — and does your vendor contract distinguish them?
• For open-source components, have licence terms and any pending patents been reviewed (Altair and Siemens funding plus the CADFit patent filing are the relevant markers here)?
• Which geometry or design classes are explicitly out of scope, and what share of your backlog do they represent?
• Is integration coverage real today, or promised — a plugin “in progress” is not a workflow?
• Do you have a baseline: cycle time, review hours, respin count and cost, measured before the pilot?
• Can the agent’s reasoning be inspected when it disagrees with an engineer?

Common misconceptions

• Open-source means production-ready. CADFit is free for non-commercial work and has no plugins for common CAD software yet; availability and deployability are different claims.
• Agentic means unattended. Celus requires the engineer to approve the plan before it is carried out; ChipAgents keeps critical judgment calls with humans.
• Verification is the low-value end. Circuit Mind’s AVE targets board respins — the expensive, late-caught class of error — which is often the fastest payback in a hardware organisation.
• Domain-specific models displace the toolchain. Cadence positions Maestro as models plus AI Super Agents that drive existing EDA and SDA tools.
• A speed statistic is an ROI estimate. The factor of three improvement Nehme reports concerns sketch selection inside the loop, not end-to-end reconstruction cost.

FAQ

Can CADFit be used commercially right now?

The project is open-source and free for non-commercial work. Initial research was funded by Altair, now part of Siemens, and Nehme and Siemens have filed for a patent for the method. Nothing in the announcement establishes commercial terms, so that is a separate conversation with the parties involved.

Which announcement is the safest first deployment?

The verification-shaped one. AVE checks PCB schematics against component datasheets, application notes and an organisation’s own design review checklists, so every finding is traceable to a document your team already trusts.

Do these agents replace engineers?

Not as described. Engineers define objectives, constraints and critical judgment calls; agents carry execution forward and operate the tools. The reallocation is in who spends time on drafting versus review.

What happens when a design falls outside supported operations?

You hit the documented boundary. Loft-based and sweep-based shapes are the current weak point — “the math will struggle” — and a revised version addressing them is being developed.

How should we read the “factor of three” figure?

As evidence that a small, cheap model can accelerate one step — choosing meaningful sketches — without being load-bearing. It is a workflow-design insight, not a productivity guarantee.

Action recommendations

1. Rank candidate workflows by verifiability, not by how impressive the agent’s reasoning looks: schematics-against-datasheets first, open-ended generation last.
2. Instrument the pilot against the failure you are trying to avoid — respin count and cost for verification tools, reconstruction hours for mesh-to-CAD work.
3. Run CADFit in the non-commercial lane against your own hardest geometry set, and log which shapes fall outside extrusion, revolution, fillet, chamfer, union and cut coverage.
4. Copy the approval-gate pattern used by Celus into any agent you build in-house: plan, present, approve, execute.
5. Separate “operates our existing tools” claims from “reasons about our design” claims when comparing ChipAgents and Cadence Maestro with a generic assistant.
6. Set a re-evaluation date now, tied to the stated roadmaps: a next CADFit version, the Siemens NX plugin, and the transition of Celus out of beta.

The through-line of this round of announcements is not model sophistication but position: agentic software is being inserted into the stages where engineers once spent their slowest hours, with human approval deliberately retained at the boundary of responsibility. The practical next step for a digital owner is narrow and cheap — take one artefact your team reviews every day, one source of truth it can be checked against, and one measurable failure, then run a single gated pilot. IT Insights is tracking follow-ups in this space; questions can go to contact@itinsights.com.