Fly, Agent, Fly

Exponential Operator is a 100% human-written newsletter by Sawyer Middeleer, founder of Revi Systems. If you were forwarded this email, join some of the sharpest AI operators by subscribing here:

I spend a lot of time talking about operations and AI flywheels.

I shared my guide to building an agentic context system a couple weeks ago, but that's just one piece of the puzzle. This time, I'm going to walk through a concrete example of an actual 'AI flywheel' that runs everything from internal admin to client delivery for Revi Systems.

I'm going to break it down how it works so that you can design a version for yourself. It's going to get a little technical because the way I run ops is technical. But the concepts will translate no matter what tools you use.

What is an AI Flywheel?

The kind of flywheel I'm talking about is a system that improves itself and delivers increasingly good outcomes over time as you operate your business.

A good example is customer delivery and new customer acquisition: As you deliver projects to customers, you discover beneath-the-surface challenges and needs, which can be addressed through new features or offerings, and in turn will help you upsell and land new customers more successfully.

It's powerful but it's not a new insight. B2B companies have been exploiting some form of this flywheel forever.

But before AI, it was a slow game of telephone, constantly held up by humans with limited bandwidth and competing priorities. You'd have to rely on someone to identify useful customer insights, communicate them to the engineers, who would translate them into requirements, prioritize them on a roadmap, and finally build the features. The flywheel could, and often did, break down at each step.

AI agents make this flywheel spin faster. They ingest a ton of data, so they can extract more insights and do useful things with them, mostly autonomously. No prioritization meetings or telephone needed. These tools don't replace human judgement, but they give you a lot more throughput capacity and (theoretically) better outcomes in the end.

Revi Systems’ core business is building a flywheel just like this. I get brought in as a fractional operator to fix burning ops issues and run core rhythms of business. Then I discover the deeper patterns that translate across clients and build new capabilities into my service. If I do my job right, I deliver better & faster for every future client.

Over the last six months, I've gradually made this flywheel more AI-native. Most of my efforts have been in bringing autonomous agents into 3 main processes:

Everything is built on two pieces: a knowledge base and an agentic automation engine that serve as our operational iron man suit. The knowledge base is the primary source of truth for everything - clients, projects, deliverables, decisions. The automation library is the engine that owns the automated rhythms which capture, synthesize, integrate, and deploy information within the business.

This isn't some "I have 9001" agents automating my business BS. It's a workflow orchestration engine driving agents - a few Hermes agents and Factory droids - that do real work. I've been using it every day for 5 months and it's constantly evolving.

By the way if you want to skip the writeup and go deeper into the technical side, I've open sourced my knowledge base and automation library architecture

The Knowledge Base

All the fancy workflows are pointless if your agents don't know how to produce good output. I've written about this before so I won't return to old ground. TLDR - the Revi Systems knowledge base provides any person or AI agent with all of the resources and context required to understand my business, the projects I've been hired to do, and how work gets done.

At the end of the day, it's really just a file system organized in a common sense way. What turns it from a basic folder on my computer into a knowledge base are the systems around it that

There are A TON of context management tools out there. I wouldn't pay for any of them, personally. Mine is built on Obsidian (free) and stored in GitHub (also free). Obsidian is nice because it works with markdown, which agents prefer, and it has a good human-facing user interface. GitHub stores my context in the cloud and makes it easy for any number of humans and AIs to contribute to the knowledge base without overwriting each other or getting confused about whether "proposal-v4.md" or "proposal-john-revisions.md" is newest.

Other teams might use Google Drive or Notion to accomplish a similar job. Those tools are fine, but I like having more flexibility and ownership of my context.

Every file is linked with other relevant notes, tracks ownership, edits, and categorization. I have linting rules that are a combination of deterministic (is file metadata correct?) and AI judgement-based (is the new content redundant or contradictory?).

Lastly, files are assigned owners and certain sections are restricted to human-only access. Permissions and governance are areas where build-vs-buy tips in favor of buy. Building in authentication and file access rules has been a major pain compared to what most software tools come with.

This knowledge base serves as the launchpad for every AI agent I run. Client asks for a project update? Agents review project trackers and client deliverables, then draft a response for me using context in the knowledge base. Need to write a proposal? Agents mine call notes and contact records and draft the document. Onboarding with a new client? Agents audit context, analyze existing operational processes, and recommend a prioritized action plan for me to own.

So not only does the knowledge base hold information about the state of the business, it also contains the instructions, skills and templates that enable my AI systems to perform well when prompted.

The Agentic Automation Engine

While my knowledge base is designed to be used by humans and AI, the automation engine is purely for the machines.

It's a little like n8n or Zapier in that it's orchestrated around triggers and scheduled tasks. But unlike those platforms, I don't have to build out workflows. All I have to do is give my agent a goal and a definition of success. Maybe it's something basic like 1) Summarize the meeting notes 2) Make sure they fit a certain structure 3) Save them to a file.

It's an agent that's usually using a skill plus additional context, activated on a trigger such as a schedule or an event (maybe a meeting is finished or an email is received) in order to do its work. That's pretty similar to a workflow engine but then there's another level, which is a bit more agentic.

The engine runs on top of a kanban project board that has tasks such as fixing a bug, researching something, or updating my site to optimize SEO. These tasks are added to the board based on all sorts of triggers. Sometimes I'll just dictate into my phone as I have an idea and that will add a task to the board. To-dos for my meetings are automatically added to the board as tasks. When something in one of my tools errors out, that error will automatically get converted into a task on my board.

The automation engine coordinates agents that automatically pick up these tasks, complete them according to the instructions in them and according to the specifications, and then leave them for me to review. Nothing gets finalized or pushed public or sent to a client without my explicit approval.

Agents are instructed to pick up tasks that are assigned to them and ones that are the best fit for their skill set. Admin tasks, for example, are often assigned to Hermes and other things like updating notes or bug fixes are assigned to a Factory droid running GLM 5.3. Perhaps more complex tasks like client deliverables are often sent to Fable or Astra, which can then coordinate smaller executor models.

You can see in this case there's not too much difference from how I might assign tasks to virtual assistants or junior employees who are expected to understand the work that I'm doing. It's a little bit more flexible and more agentic than your standard workflow engine.

But that doesn't mean that this flywheel is fully automated...

Agents can add tasks to the board themselves, but Jev helps sort which ones an agent can handle and which ones need a person. When an agent is missing context or gets stuck, it’s instructed to ask me. That catches a lot of wasted effort, though I still have to make the call sometimes.

There are deterministically enforced guardrails, too. Any actions that would update my knowledge base or modify any internal tools need explicit approval from me (in the form of merging a pull request). Agents can create a lot of slop when you aren't watching them and I occasionally need to send their work back for revision or finish it myself when it's not up to standard.

I've estimated that me knowledge base combined with automation engine enable me to currently take on about 50-70% more work than I could without them. But they're not perfect and as I write this in September 2026, there are various barriers that prevent me from pushing +100% or more.

What I'm really interested in now is figuring out how to maximize the value of the personal agent systems that have kind of taken over. There's GrokBot, Hermes, OpenClaw the OG, and now Dot from OpenAI.

This is a form factor that I really love for non-technical technical use cases because these agents are designed to be generalists. The platforms often give them access to tools and an architecture that makes them really good at remembering things about you and acting or behaving proactively.

In my experiments, I still see areas where there's still a long way to go before they can graduate from being a really cool demo to a productive agent. A lot of what I've done here is an effort to make these things truly productive, and so far I'd say partial success there.

That’s all for today. Thanks for reading!

江湖再见🫡

-Sawyer