I’ve spent most of my career building information systems for energy. Some are as simple a system as a model adding or multiplying X and Y to get Z. Some are as complex as what Halcyon is building now, a multi-source, always-on U.S. energy information system that brings structure to chaos. This is what I do, and it’s also what I think about all the time.

Way back in November 2025, developer Peter Steinberger launched the OpenClaw autonomous AI agent. OpenClaw, for those who aren’t fully lobster-pilled, is effectively a collection of independent bots that execute tasks using large language models. Humans communicate with it via messaging platforms, rather than a graphical user interface. In theory, an OpenClaw assistant can do all sorts of stuff that a person either cannot do or does not want to do for lack of time or technical capability. As an idea, autonomous agents are a compelling challenge to existing enterprise ways of working through established CRM systems. OpenClaw’s rapid rise was also an implicit (at minimum) challenge to large incumbent software companies and to the problems that they are meant to solve.

What is more interesting, to me, are the problems that autonomy, user-defined workflows, and hard-to-audit processes do not solve, at least today. First among them is the so-called ‘system of record.’

System of record

A system of record is a business’ authoritative source of its business-relevant data. This could be information on customers, or SKUs, or HR policies, or anything that needs to be consistently held and consistently available across a workforce. It seems obvious to say, but systems of record tend to be highly secured, and they also tend to be enterprises themselves, particularly as companies grow. No one wants to have 80 different sources for one finite, if very large, set of company SKUs. By the same token, no one wants to have 80 different systems for accessing the same information across hundreds or thousands of office locations.

Coordination problems do not feel like something that autonomy can solve. If anything, agentic systems could make coordination problems and interfacing with a system of record actively worse, not better: thousands of autonomous entities bombarding systems, using their own rules, leaving (or not leaving) their own logs, and initiating actions could be a security headache. And, a legion of autonomous agents could actively institute inefficiencies, trading greater speed for less coordination.

But there is something else that autonomy does not (yet) solve within enterprises. Not the ‘system of record’ as such, but what I think of as the ‘process of record.’ That is: not just where information lives and how an enterprise interfaces with it, but how organizations make and enact decisions using information.

Process of record

Every company has its way of doing important, collaborative work. Some of it is effectively enshrined (think of Amazon’s famous six-page memos, read in silence before major meetings). Some of it is less formal, but nonetheless structured, and some of it is still personalized but bounded by a needed outcome (ask me sometime about how differently my fellow co-founder Bruce and I approach our weekly team meetings to get the exact same outcome!).

Some of these processes are very important for arriving at or enacting a company’s key decisions — think of a publicly listed company’s quarterly earnings call, or the announcement of a merger or acquisition. Underneath those templated processes are another set of templated processes: due diligence, financial modeling, a step-by-step communications and marketing plan, and so on.

Importantly, each of these can be both highly individualized to an organization while also being highly systematized within an organization. This combination is effectively the productive version of “this is how we do things here,” and when done right it is efficient, well-understood, and well-appreciated.

While I am still skeptical about fully autonomous AI agents interfacing with every aspect of company processes, I have no doubt that AI will play a huge role within processes of record. But, it will probably do so with controls that ultimately should aid rather than hamper productivity, such as token limits, regular audits, activity logs, and code reviews. There will be autonomy-related processes of record underneath a company’s processes of record, in other words.

If system of record is the what of company work, and processes of record are the how, then I think there is a third organizational and informational consideration: the through of doing work with external information and inputs. And it is this through that is the most interesting to me in the AI age of speed and customization. I think of it as a ‘system of action.’

System of action

At Halcyon, we think of our work as an essential layer in both sourcing information (akin to what is in a system of record, but in this case from unstructured-but-canonical external sources), and then transforming it into the inputs for your processes of record.

That necessitates building everything we’ve built so far: collecting and organizing every state public utility commission, the most important federal energy-related sources, and important adjacent sources such as air permits. It requires an intense focus on informational context, without which the tens of millions of machine-read pages are just text. As I described it recently,

We are transforming a continuously-updated unstructured data feed into a user-specified, unique, always-on structured information system.

And it also means building on top of that system. It means building the tools through which customers can search, query, and stay alerted about what is most important to them. It means creating rapidly updated subscription information sets that deliver precise information more frequently, and with greater granularity and traceability, than you can get elsewhere.

And it also means building on top of that system. It means building the tools through which customers can search, query, and stay alerted about what is most important to them. It means creating rapidly updated subscription information sets that deliver precise information more frequently, and with greater granularity and traceability, than you can get elsewhere.

It means combining the clarity of a system of record, with the specific capability of a process of record. Doing that well requires building something new: customization.

Our subscriptions right now allow customers to dive deeper into the structured data they include by investigating the source of each data point and interrogating its context, from the page in a document through the docket and all the way up to the state in which it was filed. That is Halcyon’s context for information — but we know that customers want to add their own context, and in doing so, create their own processes. A canonical information structure, applied to something like an integrated resource plan or a large load tariff, is the start of a process. A user-specified set of questions or information conditions makes it extensible into a customer’s workflows. Our goal in working with customers is to create something that becomes integral to their own processes of record, if not a customized process in its own right.

More to come on this front soon!

Forward deploying and the never-ending last mile

Two final thoughts on what systems and processes and actions mean in an AI age.

First, I am fascinated to see foundation labs working with investment banks and private equity funds to build new companies with the sole purpose of integrating AI into other companies. The rationale is that AI is moving so fast, and potentially so far, that only the companies at its leading edge really understand what it can do. This is about as far as can be from the OpenClaw paradigm of autonomous agentic work surging through companies. It is closer in spirit to the 1970s, as Ben Thompson laid out in a prescient 2024 essay. That was when back-end offices were digitized, but in a very structured and centralized way. Today’s version is akin to (or depending on the organization, quite precisely) the forward-deployed engineer, working closely with companies over months or even years to implement something that is not only new, but also constantly changing, and constantly in need of context.

The second is a concept from Aaron Levie: The never-ending last mile of work. AI catches up to previously incumbent tasks, and then displaces them. But in so doing, it creates higher expectations for what good looks like and for what great can be. As Levie says, “the same analysis that was once required doesn’t get by anymore. The same level of product design doesn’t cut it.”

It sounds daunting in a way, but to me it is also hopeful. Raising the quality of our customers’ work means that they will not just do more work with the same resources, but do better work too. Better planning of critical infrastructure. Better strategies for portfolio acquisition. Better assessments of market structures and tariffs. Better work in every way.