As enterprise AI systems move beyond isolated prompts and simple copilots, architecture becomes a bigger constraint than model capability.

A production AI application may need to retrieve information, call APIs, maintain state, evaluate intermediate outputs, apply business rules, invoke specialist agents, and request human approval before taking an action. Trying to manage all of that inside one continuous agent loop quickly creates problems with context, reliability, and observability.

Graph-Based AI Architecture addresses this by representing AI systems as interconnected nodes, relationships, and controlled state transitions. Instead of asking one model to manage an entire process, graph engineering makes the structure of the system explicit.

For enterprises, that distinction matters. The graph becomes a mechanism for controlling how AI reasons, retrieves knowledge, moves between tasks, and interacts with business systems.

Graph engineering puts orchestration logic into the architecture where engineers can inspect, test, and govern it. Typed edges, modular execution nodes, managed state, and independent validation are some of the key differences between graph engineering and monolithic loop architectures.

3. Structured State Management

State is where graph architecture becomes particularly useful for enterprise AI.

A monolithic agent often accumulates instructions, retrieved documents, API responses, previous reasoning, and tool results inside the same conversational context.

A graph-based system can instead maintain an explicit state object.

For example:

customer_id → request_type → retrieved_records → risk_score → approval_status

Individual nodes receive only the fields required for their task and return clearly defined outputs.

This reduces unnecessary context consumption and makes it easier to understand how information changed during execution.

State can also be checkpointed. Long-running workflows can pause for human review, recover after an infrastructure failure, or resume without restarting the entire process.

For enterprise AI solutions, these capabilities are often more valuable than giving the model greater autonomy.

4. Graph-Based Knowledge and GraphRAG

Workflow orchestration solves only part of the problem. AI also needs reliable knowledge.

Conventional RAG typically retrieves semantically similar chunks from a vector database. That works well when the answer is contained in a relatively small number of passages.

It becomes less effective when answering a question depends on relationships.

Suppose an AI agent is asked:

Which suppliers serving Product A are affected by a component manufactured at Facility B?

The answer may require traversing multiple relationships across suppliers, products, components, facilities, contracts, and dependencies.

A knowledge graph represents these connections directly.

GraphRAG can therefore supplement semantic retrieval with graph traversal. During ingestion, entities and their relationships are extracted and stored as nodes and typed edges. At query time, the system retrieves a relevant subgraph rather than relying only on similar text.

This allows AI systems to reason over connected enterprise knowledge instead of treating documents as independent chunks.

Graph structures can also support persistent semantic and episodic memory for agents, while retrieving only the relevant local subgraph rather than processing an entire knowledge base.

A Practical Graph Engineering Process

The difficult part of graph engineering is not drawing nodes and arrows. It is deciding where boundaries belong.

A practical implementation typically follows several steps.

Start with the business process, not the model. Map the decisions, data dependencies, validation requirements, and failure conditions involved.

Identify deterministic and probabilistic work. Use AI where interpretation or reasoning is required. Keep predictable operations in conventional software.

Define node contracts. Each node should have a clear responsibility, expected inputs, outputs, timeout behavior, and error handling.

Design state explicitly. Avoid passing complete conversation histories between nodes unless they are genuinely required.

Add conditional routing. Define what happens when confidence is low, validation fails, data is unavailable, or approval is necessary.

Instrument the graph. Track node latency, model usage, state changes, routing decisions, failures, retries, and cost.

These practices allow teams delivering AI/ML services to debug AI systems more like distributed software rather than treating the LLM as an opaque application.

Where Graph Architecture Actually Makes Sense

Not every AI application needs a graph.

A straightforward summarization tool or single-step classifier can usually remain simple. Adding graph infrastructure to a linear problem creates architecture without corresponding value.

Graph-Based AI Architecture becomes useful when the workflow contains multiple decision paths, dependent tasks, complex relationships, parallel processing, persistent state, or independent validation requirements.

Common enterprise examples include:

• Fraud and financial investigation
• Supply-chain dependency analysis
• Complex customer-service automation
• Cybersecurity investigation
• Scientific and engineering discovery
• Compliance workflows
• Multi-agent research systems
• Enterprise knowledge assistants

Research highlighted by MIT has also demonstrated how graph-based representation can help AI identify structural relationships across seemingly unrelated scientific concepts, showing that graphs can contribute not only to retrieval but to more sophisticated forms of relational reasoning.

Graph Engineering Is Ultimately About Control

The value of graph engineering is not that graphs somehow make an LLM more intelligent.

They make the system around the model more structured.

Models can remain probabilistic while routing, state management, validation, access controls, and escalation paths remain explicit and observable.

That is why graph-based architectures are increasingly relevant as enterprises move from AI assistants toward systems capable of participating in operational workflows.

For organizations working with AI/ML development, the architectural question should therefore extend beyond which model or agent framework to use. It should include how execution is decomposed, where knowledge relationships are represented, how state moves through the system, and where deterministic controls need to constrain AI behavior.

Graph engineering provides a practical way to answer those questions.

The next stage of enterprise AI will not be defined only by larger models. It will increasingly depend on how intelligently those models are connected to knowledge, software, other agents, and human decision-makers. 

Information on the smart speaker was found in the company’s own code base.

The lowest Apple Watch Ultra 4 price can be found at Amazon this weekend thanks to double-digit savings.

Pick up the robust Apple Watch Ultra 4 today and save $20 on select styles.

Save on the Apple Watch Ultra 4

Whether you’re looking for the translucent Ocean Band or the comfort of the Trail Loop, there are several styles that are discounted at Amazon despite launching just days ago on September 22.

If you don’t need the durability of the Ultra 4 or are looking for a smaller case size, there are also several deals in effect on the S12 in our Apple Watch Series 12 Price Guide.

We’ve included highlights across the 2026 Apple Watch line below.

Apple Watch Ultra 4 deals

• Natural Titanium Case, Translucent Gray Ocean Band, GPS + Cellular: $779.99 ($20 off)
• Black Titanium Case, Burgundy Trail Loop in M/L, GPS + Cellular: $779.99 ($20 off)
• Black Titanium Case, Burgundy Alpine Loop Band in Large, GPS + Cellular: $779.99 ($20 off)
• Natural Titanium Case with Sand Trail Loop Band in M/L, GPS + Cellular: $779.99 ($20 off)

42mm Apple Watch Series 12 discounts

46mm Apple Watch Series 12 sale prices

If you’d like to learn more about the new Ultra 4, be sure to check out our hands-on Apple Watch Ultra 4 review, which goes over the pros and cons of Apple’s rugged wearable.

Battery electric cars took 29% of new car sales in Europe in August, and across the European Union for the first eight months of the year they reached 21.7% of registrations, exactly level with petrol. Hybrids remain the largest single category on 36.6%, and the spread between Norway at 98% and Croatia at 4% is twenty-four fold.

Battery electric cars took 29% of new car sales in Europe in August, according to the International Council on Clean Transportation, InsideEVs reported. Registration data from the industry body ACEA puts the figure at 29.2% across the European Union, Britain and the EFTA countries. A year earlier it was 20.2%.

The year to date is tighter.

Across the European Union alone, battery electric reached 21.7% of registrations over the first eight months, on 1,641,333 cars, ACEA reported. Petrol reached 21.7% as well. The two are level for the year.

Diesel is close to a rounding error.

It took 7.3% of European Union registrations over the same period, down from 9.4% a year earlier. Petrol fell from 28% to its dead heat with battery electric. Battery electric itself rose from 15.8%, which is the fastest move of the three.

Hybrids remain the largest group.

They took 36.6% of European Union registrations to August, or 2,759,718 cars, with plug-in hybrids accounting for 10% of the market and 758,082 of those. That is nearly 15 points ahead of battery electric, which is bad news for 2035.

The continent does not move together.

Norway registered 98% electric from January to August, Denmark 82%, Finland 50%, Iceland 47% and Sweden 43%. Croatia managed 4%. A gap of twenty-four times sits inside a market that shares one set of emissions rules and one set of tariffs.

August was a step, not a drift.

The share sat at 25% in June and 25% in July, so August breaks a plateau rather than extending a trend. BMW Group led European manufacturers with 30% of its sales electric across the year, ten percentage points up on the same period last year. Registrations across the union rose 5.3% overall.

Germany has already turned.

Volkswagen is now taking more orders for electric cars than for petrol and diesel at home, and has cancelled extra shifts at Wolfsburg while adding them at Emden and Zwickau. Germany is Europe’s largest car market, and the four models driving that shift are all built in Spain. The continent is buying faster than its factories were arranged to supply.

September brought with it an avalanche of new tech. Between Apple’s Surprise and shine, Meta Connect, IFA and Dyson Unveiled, we were inundated with gadget announcements. If you’ve managed to keep up with it all, well done to you. But for anyone seeking a recap, you’ve come to the right place.

Below, I’ve rounded up all the hottest gadget announcements from September, ready for you to browse. You’ll find the most exciting of the Apple announcements, headed up by the brand’s long-awaited first foldable phone. There’s also Meta’s camera-free smart specs, which may or may not have been created in response to the unwelcome ‘pervert specs’ moniker that’s been circulating of late. Dyson’s first ever toothbrush is there, too, in all it’s wildly over-engineered glory. And that’s just the start of it.

On Oct. 1, Google plans to hitch a ride on a SpaceX Transporter 18 rideshare flight, entrusting the mission with a test satellite it hopes will prove durable enough to start doing some light AI work while orbiting in space.

This isn’t exactly putting a data center in space — a much-hyped concept that AI entrepreneurs like Elon Musk have been touting over the last year —but it’s not not doing that. The satellite, which The New York Times reports is called “MWP” and which it describes as about the size of a refrigerator, won’t be running any huge medical research studies or beaming back streaming data. It will be low-powered, running off solar power, and designed just to test the concept of AI computation in space. Google is running the test in partnership with the satellite imaging company Planet and as part of Google’s Project Suncatcher initiative, which was announced late last year.

In a blog post that includes videos showing some of the preparation Google’s research teams have done — such as making sure that TPUs (a type of AI chip) can withstand the vibration of launching into space — the company characterizes this test as a small early step in a long-term moonshot: “exploring whether space could one day host scalable machine learning infrastructure.”

In a post on X, Google CEO asked, “Can our TPUs survive and operate in space? Well, we’re going to find out.”

Finding energy for AI in space

Google explained in the post that one reason it’s going to try crunching AI data in space involves energy. Satellites can access eight times more solar power than earthbound structures, providing a potentially huge energy resource for space data centers, which could bypass the backlash facing data centers across the country.

But that could be easier said than done. Companies like Google, SpaceXAI and Nvidia are all looking to deploy their AI technology beyond the Earth’s surface, but face some basic physical challenges: Any data center satellites would need to be built to withstand radiation, extreme temperatures and the physical stress of getting there. They’d also need to avoid dangers from an increasingly crowded orbital sky filled with other satellites and debris.

Google seems to be moving fast, with plans to continue advancing Project Suncatcher with some tangible results in the next few years; that’s assuming its October test is successful, and things keep moving forward. One of its next milestone goals, for 2027, would be to start testing high-bandwidth laser links after launching two more AI satellites with Planet. Communicating with lasers, Beals said in his Google post, is a lot harder than it sounds. “Maintaining the necessary connection requires extraordinary precision, similar to hitting a coin-size target from miles away while both points are in motion,” he said.

A data center backlash solution?

One reason space data centers sound so attractive to tech companies is a massive public backlash against building new data centers or completing ones that are in the process of being built. A pro-data center PR campaign backed by tech giants may not be enough to counter negative sentiment and protests aimed at data centers that help enable data-hungry AI technology.

Protests have sprung up across the country over energy and water use, pollution and the potential physical effects on those who live near data centers, among other issues. It could also be a point of voter discontent heading into this year’s midterm elections.

Putting data centers in space, presumably, could sidestep the issue of communities not wanting data centers nearby. But it will be costly, and there’s no guarantee that the business model of data centers in space will pay off enough for tech companies to make it viable, said Mahdi Eslamimehr, executive vice president at Quandary Peak Research.

“Proving that chips work in orbit is not proving a business model,” Eslamimehr said, and it won’t lower anyone’s electricity or water bills anytime soon or replace current data center projects, even if the business of data in space proves workable. “I welcome the experiments, but I expect everyday AI services to remain overwhelmingly dependent on terrestrial facilities in 2027 and 2028,” he said.

“A promise about 2035 does not pay an electricity bill in 2026,” he said.

Omar Gallaga has covered technology, digital culture and other topics for outlets including CNET, NPR, WIRED, Texas Monthly, MSNBC, Consumer Reports, The Washington Post, the Los Angeles Times, The Atlantic and the Austin American-Statesman, where he was a longtime tech reporter, editor and podcaster. He lives in the Texas Hill Country.

See full bio

ProfessorBoots has tons of experience 3D-printing remote excavators that can climb boxes, scoop dirt, and crush cans. His latest version parks a three-axis camera in the cab and wires a chair that copies the layout of a full size machine, so a look to the left turns the camera left and a pair of pedals drive the tracks. The result is a backyard digger you operate as if you were sitting in a Caterpillar, only without diesel and without twenty acres.

CADDXFPV sent him a GM3 combo in February, a 46-gram gimbal that tracked head movements using Avatar HD goggles and held a 19-millimeter camera by pitch, roll, and yaw. He installed the first unit on a wheel loader and tested the Goggles L, which cost $199. At first, gazing around simply turning his head seemed bizarre in the greatest possible ways. The articulated steering on that loader shifted the vision in ways that his head did not, and the original cab simply would not fit over the camera. He let the idea lay for a time until that larger, tougher excavator was ready.

Sale

Bambu Lab A1 mini 3D Printer + LED Lamp Kit, Ready-to-Use for Beginners

• A1 mini + LED Lamp Kit for Creative Light Projects: Bring your ideas to life with the included LED Lamp Kit. Simply print compatible lamp models and…
• The Perfect 3D Printer for Beginners: A1 mini 3D Printer is designed to make 3D printing easy from day one with automatic calibration, simple setup…
• Experience the Bambu Lab Ecosystem: Access MakerWorld’s huge library of ready-to-print models, manage prints through the Bambu Handy app, and enjoy…

Fitting the gimbal required a significant expansion of the cab. He imported the camera model into CAD, cloned the previous cab, removed the most of the floor, and created a new frame with a hint of design. Because the video transmitter warms up enough to potentially burn skin, he added a printed enclosure to keep a fan above it rather than beneath the lens. An aesthetic bar across the front was supposed to clip the camera when tilted, but it didn’t work out, so he ripped it out, checked clearance, and reprinted the part. The FPV gear draws power from the same battery as the excavator and operates independently, thus the same camera works on both the ESP32-robotic and the conventional radio versions.

First-person view made some jobs much easier, like being able to glance down and see the pins instead of guessing from across the yard while coupling a bucket or thumb certainly sped things up. In reality, he drove the machine to a neighbor’s card shop with a $10 payment and a handwritten message simply because he could. Nobody answered the door on his first visit, but a later knock ultimately got him in. Many RC builders do not get to compose sentences like this.

Even with the handheld transmitter, he felt like he was playing with a toy. He desired swing and dipper on the left stick, boom, bucket, and thumb on the right, and footprints on his feet. Early on in the project, two ESP32 boards communicating via ESP-NOW accomplished this without the need to hack open a radio. Early joysticks were simply assembled from a game controller with some printed extensions and springs purchased from Harbor Freight. The joysticks were dead in the middle, making fine motions clumsy, and the springs were simply too soft overall. He rebuilt the sticks with 10K potentiometers and bearings, superglued down the handles, and replaced the soft springs with firmer ones. The code now samples the resting point at boot-up and sets it as the center, so worn pots and printed parts that shift slightly over time do not cause problems.

Pedals began with a basic potentiometer pivot that allowed for proportional speed control. What happened next was a bit of a hack: both feet lifted the entire base off the ground. Then he added a printed lock, moved the spring supports further out to enhance resistance, and replaced the soldered wires with tidy plugs, allowing the entire pedal unit to be easily removed from the chair. It turns out that the signal dropouts he was seeing were caused by a residual board setting from the Bluepad32 days. Swapping the sketch to a standard ESP32 development board significantly reduced the lag. Later, a small custom PCB arrived and resolved the tangled wiring by mounting everything on neatly organized terminal blocks. He also added a fuse, battery mounts, and a power switch to prevent anything from being soldered directly to the chip.

The chair’s PLA began warping in the summer heat, pulling a spring plate completely out of shape. So he recreated the chair pieces in PETG but preserved the existing pots and bearings. Not long after, he created a desk-mounted version with the same controls, but without the office chair. Yeah, it turns out you can build the entire thing for about $60, assuming you already have an excavator and a chair. To be honest, he prefers the cheaper Goggles L more than the costly pair he purchased later. The pricey ones leaked light and required some printed shims to fix.

The ESP-NOW has a somewhat lesser range than a typical radio, but it nevertheless managed to control the machine in the yard from indoors. The problem was that it only worked for about thirty minutes before becoming clogged with dirt. His brother tested the chair and reported that it felt similar to a real excavator. Later, he used the same goggles to check under a porch to see if a neighbor’s cat was hiding; he had to tape a headlamp to the goggles to get a good look. Voltage sag killed the feed on the first pass, however fresh batteries fixed it no trouble. Unfortunately, that particular location wound up on a shortlist alongside crawl spaces and wasp nests.