"This loop becomes the new IP of the firm." Satya Nadella, June 2026
Last week the community led by our engineering team at DeepTempo shipped Vigil 1.0. A week before that I wrote about where the learning lives in an AI SOC. This post steps back from both to consider the broader pattern: where learning lives, and how it relates to the traditional pillars of enterprise IT, the systems of record.
What follows is my attempt to define Systems of Learning as a category and to say what it means for cyber.
[IMAGE 1: record vs. learning. Caption: The record is gathered and shared. The learning is consumed and improved.]
Systems of record, under strain
An enormous amount of work went into the systems of record that we all depend upon. The ERP, the CRM, the SIEM, the threat intelligence platform, the ticketing system. They earned their place by being the one location where a certain kind of ground truth is gathered, reconciled and shared. They are imperfect in ways anyone who has reconciled a CMDB can describe at length, and they are trusted anyway, because the alternative is a dozen spreadsheets.
Agents stress systems of record from both sides. Agents are now users of the record, pulling information far faster than humans. Agents can also be prolific authors, creating volumes of new rows, cases, notes and dispositions that the schemas were never sized for.
Many incumbent vendors have noticed. Most are shipping agents that run inside their own environment. The forward-looking ones are also opening the system of record itself to agents that live elsewhere. Salesforce calls its version Headless 360: everything on the platform is exposed as an API, an MCP tool or a CLI command. That seems a gutsy and forward-looking approach to capturing the rapidly increasing usage by enterprise agents.
Systems of Learning
Just as a system of record earns its place as the location where ground truth is gathered and shared, a System of Learning earns its place as the location where the context held in many systems of record, and other context, is consumed by intelligence, and where the system improves and stores improvements to the learning architecture.
In a recent blog on how AI SOCs can learn and retain their learning, I borrowed from Chen-Yu Lee's Berkeley lecture, From Goldfish to Self-Improving Agents, to describe the five places that learning can live outside a model's weights: memory, lessons, skills, environments and the harness itself. A system of learning is the thing that holds those five, keeps them readable by the people accountable for the outcome, and feeds them with judgment gained from operations and experience.
[IMAGE 2: five places. Caption: Five places learning lives outside the weights, after Chen-Yu Lee.]
The firm's advantage is in the learning
My observation, drawn from watching large organizations work through this over the last year, plus a lot of experience of my own over a few decades, is that while a firm can continue to outsource its systems of record, it must own and control its own system of learning.
The reason is simple. The system of learning can be a major source of differentiation, and each enterprise's unique capabilities will be reflected in it. It cannot safely be outsourced, because whoever owns the loop owns the compounding improvements and the differentiation itself.
Legal is the clearest early example. Harvey and Legora are not winning because they have unique access to particular LLMs. It seems that they are winning because they are becoming the system of learning inside each firm: the place where the firm's precedents, playbooks and corrections accumulate and where the next matter benefits from the last one.
What the System of Learning means for cyber
Some have argued that cyber is the next domain Anthropic and its peers will simply take. In a narrow slice of the market they are right.
Vulnerability discovery is close to a pure capability problem: Claude Mythos Preview found thousands of high-severity zero days across major operating systems and browsers, and Project Glasswing is the sensible response. The model's intelligence is nearly sufficient to resolve the use case. Context is almost beside the point.
In most of cyber, context is the whole point. Which service account talks to which database at three in the morning. Which subnet never touches the internet. Which detection was tuned down last quarter and why. Customers cannot and will not share their security posture publicly, and that posture is exactly the context required to make good decisions at scale. No frontier lab can ever hold it. The firm must.
So I expect that cyber use cases will, over time, be managed through a system of learning that looks most like a cyber fusion center: the place where investigation, hunting, detection coverage, response and, increasingly, agents hunting agents are supported, merged and transformed into one improving operation. The relevant systems of record remain in place. OpenCTI for threat intelligence, Splunk or Wazuh for the SIEM, the ticketing system for the ticketing. Logically, though, the intelligence and the improving system for cyber operations must be independent of them, and must be owned, trained, educated, monitored and scaled by each organization.
I wrote in September about the uncanny valley between an agentic SOC and an autonomous one, where a SOC that merely pedals faster earns a faster ticket queue and loses the trust of the people watching it. The way across that valley is the same architecture I am describing here. Own the loop. Keep the declarations under change control. Let the record be the record.
Where Vigil fits
We built Vigil to be this system of learning for cyber, and 1.0 makes this pretty clear.
[IMAGE 3: Vigil as the system of learning for cyber]
Vigil sits between the record and the people, consumes the SIEM, the TIP, the telemetry and, optionally, LogLM's compound detections, and keeps its memory, lessons, skills, rehearsal environments and harness edits where an operator can read, review and delete them. It is Apache 2.0 because you cannot own a loop you cannot inspect.
I would welcome disagreement on any of this. Is "system of learning" the right name? Where would you draw the outsourcing line in your own stack?
I am working on a longer whitepaper and would really appreciate any thoughts and collaborators. Does the System of Learning help make sense of the current adoption of AI and what to retain versus what to outsource? How would you explain and expand upon the concept?
