Has Agent Pricing Gotten Any Better?
A little over a year ago a client challenged me to consider what agent pricing would look like in a couple of years. After digging in for a while, I came to see a number of very real customer and vendor issues — and in general an area ripe for change. I wrote a three-part series about it with some predictions for the next 12 months. Since then, there has been a very steady stream of vendor pricing changes (including from the aforementioned client). But the only consistent theme from every briefing was that, while a given pricing update was new, the broader effort to refine pricing was still a work in progress, so we should expect still more changes. This prompted me to look at the situation again to see how my predictions fared (TL;DR — not too shabby) and to frame what I think is coming next. I also wanted to offer some specific advice for enterprises as they consider new contracts and renewals. Last week, part 1, covering my previous predictions and the current state of agentic pricing, went live on Forbes. This week, we publish part 2, which contains my new predictions and advice for enterprises.

Are Knowledge Graphs a Product Category
AWS announced Amazon Context at its AWS Summit event earlier this summer. Amazon Context is a knowledge-graph-as-a-service for enterprises to map metadata and their relationships across multiple structured and unstructured data stores. It’s an interesting concept, and I am hearing that other products like this should reach the market before the end of the year. Knowledge graphs are remarkably powerful tools for large and diverse datasets, and they have performed particularly well as a context map for agents. If you have heard the term “second brain” used for Claude or OpenAI, that is a personal knowledge graph.

So, while it is a powerful technology, is it a standalone product category? Until now, knowledge graphs have largely been either embedded into a solution or built bespoke for a specific data exercise. The most popular embedded solution is the Obsidian note-taking app, which is the second brain of choice for both Mike Leone and me. On the bespoke side, one company that has mastered building knowledge graphs is espalier.ai, which introduced me to the concept about six years ago. But the idea of a knowledge graph as its own out-of-the-box solution is new. I am not 100% sure it will take hold, though I suspect it will become a common data platform feature like vector databases, as Mike pointed out last week in his Field Note on DynamoDB.

SaaS-Based Agent Control Planes Are Really Good
Last week I was able to take a deep dive on Mulesoft’s Agent Fabric control plane, which is part of Salesforce’s Agentforce platform. What stood out most for me was not the feature set but how SaaS companies are adapting products and points of view as agents continue to threaten them. For instance, SaaS in general is well known for its walled-garden approach to integration, and in fact when Salesforce acquired Mulesoft in 2018, many thought it would somehow limit Mulesoft’s capabilities. But Salesforce has done a nice job of not messing with the Mulesoft approach, and that really shines with Agent Fabric. For instance, while many control planes have cross-platform agent registries, Agent Fabric automates agent discovery with easy-to-deploy scanners that auto-register newly created agents in AWS or Azure. I was also impressed with the policy management capabilities implemented in the OmniGateway. If there is a downside to Agent Fabric, it may be some of the observability and FinOps capabilities, though more of that functionality is on the roadmap, and I am hoping to hear more at Dreamforce later this year. That said, even for those who are not Salesforce customers, Agent Fabric looks like a competitive and in some ways even leading agent control plane. So, once again, Salesforce — and, as I wrote earlier this year, SAP — are not being passive about the threat of the agentic future. In fact, SAP also announced last week that its control plane is now generally available.

Has Agent Pricing Gotten Any Better?
A little over a year ago a client challenged me to consider what agent pricing would look like in a couple of years. After digging in for a while, I came to see a number of very real customer and vendor issues — and in general an area ripe for change. I wrote a three-part series about it with some predictions for the next 12 months. Since then, there has been a very steady stream of vendor pricing changes (including from the aforementioned client). But the only consistent theme from every briefing was that, while a given pricing update was new, the broader effort to refine pricing was still a work in progress, so we should expect still more changes. This prompted me to look at the situation again to see how my predictions fared (TL;DR — not too shabby) and to frame what I think is coming next. I also wanted to offer some specific advice for enterprises as they consider new contracts and renewals. Last week, part 1, covering my previous predictions and the current state of agentic pricing, went live on Forbes. This week, we publish part 2, which contains my new predictions and advice for enterprises.

Are Knowledge Graphs a Product Category
AWS announced Amazon Context at its AWS Summit event earlier this summer. Amazon Context is a knowledge-graph-as-a-service for enterprises to map metadata and their relationships across multiple structured and unstructured data stores. It’s an interesting concept, and I am hearing that other products like this should reach the market before the end of the year. Knowledge graphs are remarkably powerful tools for large and diverse datasets, and they have performed particularly well as a context map for agents. If you have heard the term “second brain” used for Claude or OpenAI, that is a personal knowledge graph.

So, while it is a powerful technology, is it a standalone product category? Until now, knowledge graphs have largely been either embedded into a solution or built bespoke for a specific data exercise. The most popular embedded solution is the Obsidian note-taking app, which is the second brain of choice for both Mike Leone and me. On the bespoke side, one company that has mastered building knowledge graphs is espalier.ai, which introduced me to the concept about six years ago. But the idea of a knowledge graph as its own out-of-the-box solution is new. I am not 100% sure it will take hold, though I suspect it will become a common data platform feature like vector databases, as Mike pointed out last week in his Field Note on DynamoDB.

SaaS-Based Agent Control Planes Are Really Good
Last week I was able to take a deep dive on Mulesoft’s Agent Fabric control plane, which is part of Salesforce’s Agentforce platform. What stood out most for me was not the feature set but how SaaS companies are adapting products and points of view as agents continue to threaten them. For instance, SaaS in general is well known for its walled-garden approach to integration, and in fact when Salesforce acquired Mulesoft in 2018, many thought it would somehow limit Mulesoft’s capabilities. But Salesforce has done a nice job of not messing with the Mulesoft approach, and that really shines with Agent Fabric. For instance, while many control planes have cross-platform agent registries, Agent Fabric automates agent discovery with easy-to-deploy scanners that auto-register newly created agents in AWS or Azure. I was also impressed with the policy management capabilities implemented in the OmniGateway. If there is a downside to Agent Fabric, it may be some of the observability and FinOps capabilities, though more of that functionality is on the roadmap, and I am hoping to hear more at Dreamforce later this year. That said, even for those who are not Salesforce customers, Agent Fabric looks like a competitive and in some ways even leading agent control plane. So, once again, Salesforce — and, as I wrote earlier this year, SAP — are not being passive about the threat of the agentic future. In fact, SAP also announced last week that its control plane is now generally available.

The big AI security incidents get reported, but what a team handles quietly inside its own walls never gets out, and neither do the near misses. Nobody aggregates what does get reported, either, so one control failure affecting three companies usually reads as three unrelated stories. Beyond that, open frameworks are multiplying, and each carries its own vocabulary for classifying risk.

SAFE, short for Shared AI Findings Exchange, is the Open Secure AI Alliance’s proposal for doing that aggregating. It’s currently a Request for Comments through the Linux Foundation, so nothing is adopted and nothing binds anyone. Reporting is confidential, so a company can report an incident it never disclosed publicly, and the recommendations that come out are public. SAFE only covers incidents after the fact, which I’d argue is the right limit. No alliance is going to stop an incident inside somebody else’s lab. An exchange is only worth what members file into it, and there’s no deadline and no enforcement behind it. If you’re buying security tooling, ask the vendor whether it intends to report. Most won’t have thought about it.

Red Hat helped draft SAFE and put out an open-source project of its own alongside it. “asago” (which stands for AI Safety and Governance Orchestration, though the acronym is in lowercase) turns a written AI policy into a tested, deployable agent without a full-time safety team. It reads the policy, then maps each risk it finds to the IBM Risk Atlas, a published catalog of AI risks. From there it generates test scenarios, runs red-teaming, and proposes guardrails. It’s early upstream code at this point, and Red Hat says the architecture will change, so nobody is deploying it yet. Red Hat contributes the orchestration and IBM supplies the vocabulary. It’s worth noting that whoever writes the risk taxonomy shapes how other tools label what they find, and swapping a taxonomy out later is far harder than swapping the tooling.

In AI, storage vendors stopped competing on the array a while ago. Two questions took over: How much data can you reach when it sits on somebody else’s hardware? And how much work can you do without moving it?

Now VAST’s AI OS 5.5 is generally available, putting container orchestration, an expanded query engine, and a native vector index inside the VAST cluster. The query engine picked up regression and correlation, so it does real statistical work now. VAST spent most of its life selling fast file access to GPU clusters. Processing data that lives in storage normally means buying separate compute servers and moving the data over to them. Some of that work can now happen inside the VAST cluster, so there’s less data in flight and fewer servers doing the processing. VAST has built out an unusually broad stack, spanning an operating system, a database, a query engine, a compute layer, an agent engine, and a vector index. VAST hasn’t said how all of this gets sold or priced yet, and that’s the harder part of shipping this much this quickly.

NetApp has been moving at a similar pace, with a second acquisition in three weeks that reaches data on other vendors’ hardware. The newer acquisition, JetStream Software, handles disaster recovery and migration for VMware environments, wherever that storage happens to sit, and the recovery copy lands on Azure NetApp Files. NetApp’s own data protection has needed NetApp storage on both ends; JetStream covers VMware on somebody else’s arrays. For an AI program, governing the data matters more than owning the hardware under it. Buying into VMware estates that NetApp doesn’t currently touch widens what it can govern. Terms of the deal weren’t disclosed. The earlier acquisition, DataPelago, added processing where the data already sits, and this one reaches a customer base NetApp hasn’t served.

Buying a vector database used to be a separate line item. Amazon DynamoDB now has vector search built in, generally available across every commercial region and GovCloud, with no public preview. (If you want to dig into the details, I wrote a Field Note about this last week.) The expensive part is the pipeline copying data from the system of record into a second store, and that copy goes away. Adding AI to an existing application usually stalls over who is allowed to see what, and having one datastore means just one set of access rules to argue about. Vectors carry their own charges for writes, searches, and index storage. Dropping the second store helps teams whose data already sits in DynamoDB, and that’s a very large group.

ThoughtSpot has announced AgentSpot, which generates a working application from a description, hosts it, and enforces each viewer’s data permissions on every load. It also runs agents against live data and turns written playbooks into scheduled workflows. Plenty of tools build an app from a prompt now, but very few inherit permissions from a governed data model, and ThoughtSpot has that layer already. ThoughtSpot also shipped the access controls and audit logging that IT procurement asks for. That puts ThoughtSpot alongside Salesforce Agentforce and Google Agentspace rather than Tableau and Power BI. Agent budgets sit with IT and operations, which is a bigger pool than the analytics teams ThoughtSpot has been selling to.

AMD Robotics
AMD disclosed a customer for its physical AI platform on July 23. Foundation Future Industries validated the Phantom MK-1 on AMD compute and plans to build the MK-2 natively around AMD parts, using Kria system-on-modules for the robot brain and FPGA-based systems for hands, sensing, and real-time control. The decision moves Foundation away from NVIDIA, whose chips and software have been the default for most humanoid programs.

My take: Last week I flagged the software story as the gap to watch, and Foundation is an early success indicator. Here is a company with robots doing paid work today, committing a production fleet to silicon that has not shipped in volume. That’s a bet on the toolchain, the partner network, and the supply commitment — not just benchmarks — which is exactly the calculation I would expect from this class of buyer. One customer is not a trend, and the MK-2 does not exist yet. But it is the right kind of first customer, and the software question now has a real product attached to it.