Cognex agreed to pay about $600 million for RealSense as Overview AI's camera sets up its own inspections

On October 6, 2026, Overview AI, a private machine-vision maker founded in 2018, began selling a camera that sets up its own inspections ¹². An agent called Sparky runs on the camera itself: it picks which of ten inspection tools a check needs, answers plain-English questions about what the line has been running, and lays out the operator screen from a description ¹³.

Overview built the agent after a decade of better vision models still left its customers a consistency gap: which tool to use and what to take from the production data came down to individual judgment ¹³. If an agent can carry that judgment onto the camera, the pressure lands on the smart-camera incumbents, Cognex·−4.57% — Cognex, down 4.57 percent as of its last quote (NASDAQ: CGNX) and Keyence, whose mid-range models Spark is pitched directly against ⁴⁹.

Setup work moved onto the camera

Overview's founders describe a decade of the same lesson ¹. "The hardest part isn't the inspection itself," says chief executive Chris Van Dyke: every new line meant another project, another specialist, another few weeks ¹. Plants have more inspection points than people qualified to set them up, says co-founder Russell Nibbelink ¹.

Sparky, the agent, carries that work onto the device ¹. Asked about a connector that must show 24 pins before it reaches the programmable logic controller (PLC), the line's control computer, it explains that the Alignment tool follows the part as it drifts and the AI Count tool checks that every pin is present ¹³.

No software key, and no subscription

Everything runs on the camera: training, inference, the PLC link, the image library and the operator screen, on a Qualcomm▼−2.65% — Qualcomm, down 2.65 percent today Dragonwing processor, sold with no software key and no subscription↗ ¹⁴. Overview's earlier cameras already trained and ran models on the device ³; the agent is the new layer. Chief technology officer Xiaoyang Kao calls Spark the first product where that onboard intelligence "can support an expert in doing their job" ¹.

The first check came from an engineer who had never used a camera

The chain already reaches a person: at Mansfield Engineered Components, an Ohio metal fabricator, quality engineer Logan Hoon, a year into the job with minimal camera experience, built a missing-washer check for a station that previously got only an operator's visual inspection ¹.

The part he expected to be hard, the alignment, "turned out to be pretty easy, because it tells you exactly what to do" ¹. The plant has since moved the camera between two lines and is scoping a second ¹.

One named deployment is the honest scale of the evidence; Overview says its cameras already run in hundreds of factories↗ ⁵.

Cognex just bought the market above the one Spark attacks

Two weeks before the launch, on September 22, 2026, Cognex↗ agreed to pay about $600 million for RealSense, the depth-camera maker spun out of Intel▼−5.15% — Intel, down 5.15 percent today, most of it in cash ⁶⁷. Cognex frames the deal as entry into robotic perception, a market it estimates passes $1.6 billion by 2030 ⁶. RealSense expects $80 million to $90 million in sales this year↗, up more than half ⁶.

Cognex's own quarters have been climbing through all of this↗ ⁸.

If self-setup spreads, incumbents must answer and plant engineers gain

If agentic setup works beyond one plant, two things follow. Stations that today get only an operator's glance become camera sales, and the setup and licensing revenue wrapped around smart cameras erodes.

Cognex and Keyence would then answer with agents of their own, or they would acquire, which would make smaller agentic camera makers plausible acquisition targets; that step is a position, not a reported deal. The pattern is familiar: an incumbent buys growth up-market while a self-configuring product enters from below, and the incumbents' next product cycle is where an answer has to appear.

Someone still has to know the part

For the people on the line, the skill to start building now is saying plainly what a good part must show, then checking the agent's answer against real parts. The agent can assemble the inspection, but someone still has to know the part, and that is learnable this month regardless of what Cognex buys next.