The flywheel - Agents that build agents
Subtitle: Verge of anything is possible
The flywheel is a business strategy and growth concept. Basically, you find a virtuous cycle that just keeps going. Like having a super-great product. Then more customers means more referrals to next customers. See how that just by itself. I suppose the most famous is the Amazon flywheel where more retailers means better selection. Better selection means better experience which brings in traffic (customers). That in turn means more retailers coming for the traffic. The Amazon is actually a double flywheel where the second flywheel is the growth-cost-lower-prices.
Agents that build agents
The simplest way to build an agent is to take on agentic coding tool. The famous ones are Claude Code, Codex and Cursor.
All you need to do with the coding agent is to say that “I want an agent for X”. Then you watch the code generating agent do the job. You will get clarifying questions. You will likely need to provide a few technical answers on how you want the agent to be built. However, the answers are starting to be quite understandable. If you don’t know the technical answers, then you work with your code generating agent to figure those things out.
I so know this still needs software architect / software engineer knowledge. Few (if any) companies have yet created the necessary context and scaffolding to just supply all agent creators the guardrail on what, where and how.
But back to building. You have prompted for the goal. You have answered questions. The code generating agent (Claude Code, Codex, …) will then make everything happen. It will make the agent definitions. It may or may not make tests. If it doesn’t, then just ask it to test the agent out. The smart ones will ask for “evals” to be created too.
(Note to self: write about Evals soon)
Let’s assume you are building something simple enough. Like an email manager agent set. That will be created within minutes. The first agentic workflows will also start working in minutes. In this example, the biggest barrier will be access to the source data (the emails). But if for instance, you have the data in your laptop, then making the first agents is really a matter of minutes.
Code generating agents build other agents. How meta?
The agent flywheel
See how this is kind of starting to get closer to a flywheel. But a flywheel really needs the feedback from success to initial drivers to really start compounding growth.
The flywheel is Agents building compounding context that helps build next agents
Agents can be built so that you all the time make it easier to build next agents. Let me give you a rudimentary example: access to sources get easier over time. You connect to your emails with the first agent use case. Then the next use cases will inherit that same knowledge and connection.
Agent examples make next agents easier to build. Code generating agents are very smart in using LLM to understand how current system works. They take on examples, generalise and then apply the next goal. The more you have correct examples, then the easier it is to make more.
There however is a deeper and a more fundamental way that the advanced users are figuring out. We are starting to create agents that are self-learning and that build the context themselves. This is surprisingly simple and doable already today.
Each agent-creating-agent gets a goal: Create X. While you create X, you must create structured learning and accumulate all needed material to do better on the next iteration.
Periodically, as part of all session to create agents: Reflect on the current session. What is working and what is not. Ensure the next sessions are better by memorising the learnings. Change your definition to improve for the next round.
The current limits to the flywheel
I have done my share of trying to make agents that learn. The cutting edge code generating agents are very close to something that can learn.
For instance, a challenge of building a simple digital twin for a complex software creation program is a doable challenge. It takes days to complete, but does complete with human guidance and a lot of LLM self-reflecting.
I feel the current limit of the LLM and code generating agents technology is the lack of memory suitable for working with complex and nuanced situations. The basic LLM remembers only so little in single chat. Once the context runs out, then the processing goes off rails. You can split your agents, then you need the memory of sorts even more. The human memory and thought-creation is excellent in making connections. LLMs are utterly unable to make connections outside of their current context window. A memory with similar properties as our thinking heads will likely be needed.
The other limit is the ceiling of meaning and correctness. Right and wrong are ultimately subjective and social constructs even in business. An agent can create whatever. It can learn to make yet more things. Any information could be created by the agent. The output and information could be right. Similar to any actions the agent took. Were the actions valid and appropriate? But rightness is only usually known after humans judge this. I know we could make LLMs that check LLMs. But ultimately, the output needs to connect with human notion of correct. And you know how long it takes for humans to agree on something.
The third ceiling is the processing. The more complex problem you give, then more computation is needed. Simple email triage can be simple and quick per email. But anything beyond gets exponentially slower. The more data, then more dimensions, the more possibilities. The more it will take for the Agent (LLM) to answer. The more it will thereafter still take to check if the agent made any sense.
The Ralph Wiggum loop of learning
People have been raving about the Claude Code Ralph Wiggum loop over the last weeks online. This loop is basically telling the LLM-based Claude Code to keep going when it is trying to call something done and dusted. Often it really not yet is. The Ralph Wiggum loop basically urges the LLM to keep going until the goal has been met.
In the Ralph Wiggum official page, it is said: “The technique is named after Ralph Wiggum from The Simpsons, embodying the philosophy of persistent iteration despite setbacks.” Those who know Ralph, know he is a bit slow and not very bright. The story goes: Chief Wiggum's (Ralphs’s dad) full advice is: "The key is persistence. Keep at it and never lose your cool." When one nut refuses to crack, Chief Wiggum takes a step back, pulls out his pistol, and shoots it, yelling, "Let that be a lesson to the rest of you nuts!"
Persistence in learning is what you need. Go check out the loop to see how simple it can be. You might be surprised.
See how building agents is only about the “doing it”. Add learning, and you see the agent flywheel.