Enterprise AI's Breadth vs Depth Problem

At Ophelia, we build infrastructure that enables software to interact with the real world. This week, Bo Brainerd sat down with AgentPhone founder Manav Modi (YC P26) to discuss the emerging stack powering autonomous consumer products. We're entering a world where software can: → Call businesses
→ Book appointments and reservations
→ Coordinate multiple parties
→ Complete purchases
→ Execute real-world requests end-to-end Everyone is talking about AI agents. Few are talking about the tools required to make them work. This is a completely new computing paradigm. The companies building the infrastructure for autonomous action today will shape how billions of real-world decisions get made tomorrow. Full conversation below.👇 https://lnkd.in/g4-4Vz59

Excited to finally share what we have worked on for the past few months. Introducing Replicorg: https://replicorg.com We started with an autonomous software factory. Software work can be investigated, built, validated, previewed and moved forward by AI agents, with many jobs running in parallel. Humans can step in whenever they want. Otherwise, the work keeps moving. Replicorg is built around a broader idea: increasingly complex company work can be understood, coordinated and executed by AI agents. The software factory is our first concrete expression of that idea.

IT product companies are entering a new phase in 2026. It's no longer about bolting AI onto existing software. It's about engineering products AI-first, deploying autonomous agents at scale, and building governance into the foundation — not as an afterthought. Swipe through the 3 shifts defining the industry right now  Which one is your organisation already living? #ITTrends2026 #AIFirst #AgenticAI #SoftwareEngineering #TechTrends

Most autonomous systems fail when the direct path is blocked. This one doesn't. I placed obstacles directly between the robot and the target. No straight line exists. The robot must find an alternative route by itself — no map, no GPS, no cloud processing. For a moment it looks stuck. Then it computes a solution and breaks through. This is not a simulator toy. The 2D visualization is a window into the control system. The same code runs on a physical robot — no retraining, no recalibration, same code, different room. No training datasets. No machine learning. No internet connection. No cables. No commands. No one in the room. One robot. Pure mathematics. Real autonomy for the physical world.

Most autonomous systems fail when the direct path is blocked. This one doesn't. I placed obstacles directly between the robot and the target. No straight line exists. The robot must find an alternative route by itself — no map, no GPS, no cloud processing. For a moment it looks stuck. Then it computes a solution and breaks through. This is not a simulator toy. The 2D visualization is a window into the control system. The same code runs on a physical robot — no retraining, no recalibration, same code, different room. No training datasets. No machine learning. No internet connection. No cables. No commands. No one in the room. One robot. Pure mathematics. Real autonomy for the physical world.

Couldn’t agree more with this… ‘The tech people are worried. Not about capex overbuild. They’re worried about the coming demand. Most people who are reading all the articles… about capex bubbles are thinking in terms of the technology we have today. Coding assistants and chatbots and user-generated artwork. But they’re not thinking about what it’s going to take to facilitate the existence of a hundred million autonomous cars and trucks, humanoid robots, automated manufacturing operations and other such things that are a little further out on the horizon. However, the people making these investments with trillions of dollars and multi-decade time horizons are thinking precisely this way.’ https://lnkd.in/e8J9cFVK

For those that decry an AI bubble/overbuild, we are just getting started with the capacity needed for a huge increase agentic AI demand for robotics, autonomous systems, transportation, manufacturing, communications, utility operation, safety, and financial systems.

Ford AI Assistant is going Live! Excited to see how our Towing customers use it to learn more about their vehicles capabilities and prep for towing trips! 🛻 Lots of care in this to make it great for our towing customers; Let me know what you think! Shout out to Ali Saleh Yao Ge Timur Gilmanov & Atticus on the AI towing support

𝗚𝗼𝗼𝗴𝗹𝗲 𝗗𝗲𝗲𝗽𝗠𝗶𝗻𝗱 𝗷𝘂𝘀𝘁 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱 𝗚𝗲𝗺𝗶𝗻𝗶 𝗥𝗼𝗯𝗼𝘁𝗶𝗰𝘀 𝟮, 𝗮 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹 𝗱𝗲𝘀𝗶𝗴𝗻𝗲𝗱 𝘁𝗼 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝗵𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺𝘀. 𝗢𝗻𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗱𝗲𝗺𝗼 𝗵𝗮𝗿𝗱𝘄𝗮𝗿𝗲 𝗽𝗮𝗿𝘁𝗻𝗲𝗿𝘀: 𝗦𝗵𝗮𝗿𝗽𝗮'𝘀 𝟮𝟮-𝗱𝗼𝗳 𝗱𝗲𝘅𝘁𝗲𝗿𝗼𝘂𝘀 𝗵𝗮𝗻𝗱. The announcement is late July. The demos ran on Apptronik's Apollo 2 with different end effectors — Sharpa's Wave hand, Inspire hands, Robotiq grippers. What Google showed is one model that can control different hardware configurations. Per Gasgoo, Sharpa's hand "served as core verification hardware" for the demonstration project. The hand handled tasks like tying trash bags, unscrewing light bulbs, and sealing ziplock bags. That is meaningful progress. The robotics stack has been fragmented — one model per robot, custom integration for every task. Google is betting on the LLM approach: foundation model plus specialized hardware. Train once, deploy across platforms. But this is not deployment-ready yet. Google's own benchmark data shows multi-finger dexterity tasks hitting 32-44% success rates. Sealing a ziplock bag fails more than half the time. Google says full-body movement "still has a way to go, with regard to speed." The model is available to early-access partners only. So the hardware has not caught up. It has gotten better. The gap between planning and execution is narrowing, but it is still there. The clearer path forward is real. One foundation model controlling multiple platforms is the right architecture. The question now is how long it takes to close that 32-44% success rate up to something a factory floor can depend on. Manufacturing, logistics, assembly — these sectors need manipulation at volume. Google just showed them the route. Walking that route is the next phase. Click here to subscribe to The Hirsch Report: https://lnkd.in/gyzanhxn Source Article: https://lnkd.in/gwrVP4pw #Robotics #AI #Google #FoundationModels #Dexterity #Automation #EmbodiedAI #HirschReport #DeepFactChecked

Google DeepMind released Gemini Robotics ER 2, a model that combines high‑fidelity video understanding with tool orchestration and multi‑robot collaboration. The system can ingest raw video streams, infer object states, and generate actionable commands for heterogeneous robot fleets in a single pass. For engineers, this means less time writing custom perception pipelines and more focus on domain logic. ER 2’s API returns structured task graphs that can be fed directly into ROS‑compatible controllers, enabling rapid prototyping of warehouse automation, inspection drones, or collaborative manufacturing cells. The model is accessible via Google Cloud AI, with pre‑built wrappers for Python and C++. Early benchmarks show a 30% reduction in latency compared to stitching separate vision and planning models. What robot coordination challenge will you tackle first with ER 2? https://lnkd.in/eKWVTT8B #Robotics #AI #MachineLearning #Developers