I’ve long believed that at the limit we are trending to a future where everyone is a founder dominating their own niche. I just didn’t expect it to be possible so soon.

With the launch of agentic workflows we are seeing AI’s collaborate to recursively solve problems. They do this by self-prompting in an infinite loop. This dovetails with my theories about 🧠Universal Learning Machines and how they must be embodied to progress. This form of self-prompting is effectively the ability to take the initiative.

It’s funny because I was theorizing something just like this for map building in the physical world — effectively Loss of Detail strategies for recursive path planning in spatial path planning. *LMs can do this with language and ideas at abstract levels.

In any case, in this post I wanted to think about the market a bit and where it is going to examine whether or not to engage with it.

The Market

It looks like it is possible to build embodied agents with agentic AI. This feels like the internet moment. Agents for everything. Let’s think through the knock on effects of this gold rush. Now that we have vision as well everything is going to change. Now that we can upload documents everything is going to change. We can effectively enter an era of content driven development. And this can be circular. The new work is the defining of outcomes and design. All implementation work is going to evaporate first. So, task definitions or management is the future of work. This is an old belief but now for the first time it is possible.

What are the knock on effects? With ChatGPT4o agentic workflow companies are going to explode. Companies are going to pick specific niches to focus on and scale out from. What are the knock on effects of this? There is going to be some convergence at some point on platform. What can scale small and become an infinite Amazon of workers who do specific tasks and can work with the specific integrations? The starting point

This has already happened and been happening with companies like Agentic, Beam, and CrewAI. Crew focuses on developer tools. Beam focuses mostly on back office through a slick UI. Agentic focuses on playtesting. LlamaIndex focuses on document querying. Where does this go? Where does value accrue and what are the defensible moats that will emerge?

Clearly a major asset is having a pool of agents who actually know how to do the specific tasks required by your company and their SOPs. I don’t believe that a hyper generic system will provide any value other then to provide lego bricks to put together company specific workflows. Once enough lego bricks come together then you can define higher level agents. This can continue on forever (potentially).

It would seem that value will accrue to whatever can assemble the most lego bricks, integrate with the most tools, and then adapt to build higher level agents recursively.

One interesting question is whether to focus on existing companies or not. Focusing on existing companies provides a specific value prop which is great. The SOPs exist. The tools are known and you can penetrate specific companies and create sticky tools. On the other hand the biggest opportunity is in all of the people who couldn’t scale themselves effectively before. They can produce their own personal workflows rather than company level workflows. They can step into a role and build a suite of agents and workflows to automate themselves away. Effectively, they define a set of agents that defines a person in a role at a company.

The difference is illustrated by a traditional team at a call center. They define jobs to be done and SOPs for those jobs and the people are interchangeable. The problem is that the level of abstraction is at the job level instead of the agent level. What I want to do at a company is unbundle the people so that I can have them automate themselves away. Everyone should become a manager. The problem is that work is not actually a collection of jobs, its a person with a set of skills that can reply to demands. The nature of work is change and so the system needs to be able to change dynamically.

This leads us to how the market will progress and what the next several moves will be.

The Next Several Moves in Agentic Workflows

The first step companies will reach for is searching for highly repetitive tasks that people do at enterprises where there is money to spend and then they will focus on creating integration lego bricks for the workers to define co-pilots and agents to facilitate those workflows. Those workflows will be created, updated, and evolved by the humans that manage the SOP for the business. Teams of AI managers will handle over-seeing what these agents do, making approvals, and so on according to the recipe.

The second step products will experiment with agents that construct or tweak these workflows because folks will be anchored on the human model of doing work. But this will be intractable. Though that is some time away. When you get away from the task level definition of work the opportunity for mistranslation explodes so I believe this will be some distance away and that for at least the next year the primary value will be just above the task level and as close to the leaf of the tree of work as possible.

The third step is flattening AI organizations into graphs of work. Since AI agents will get lost in the sauce and “drift” if you use language that is too high level it will make more sense to create an agent graph for work to propagate through. While it is true that there is the ability to recurse information gets lost in hierarchical systems where games of telephone happen. So, it will be best to stick to managing graphs of agents offering at the task level. The agent graph will be defined by a “founder” or “manager”. This agentic programmer will sits at the intersection of system design, prompt engineering, and platform integration. Work will enter on the edges, propagate cyclically through the network processing artifacts and emerge on the edge. Human oversight can now build graphs of arbitrary size. They interact with the graph by “intruding” on the network to inspect artifacts, answer questions when things get blocked, and reviewing output artifacts before they get sent out. Humans can actually exist within this graph and do work delegated by the AI around approvals, design review, and so on. Even posting jobs to sites like Mechanical Turk.

Another key problem and duplication of work that we will see is that every agentic company will need to build the same integrations with X tools. This will give rise to aggregators who provide API hubs that agents can query rather than having to build out N integrations themselves. This could even be an AWS type moment if a market leader emerges.

Another interesting opportunity is acting as an aggregator on top of other companies that specialize in say workflows for research and workflows for X. The problem is this won’t be defensible as the niches expand out into ever wider selections of workflows.

The other question is whether agentic companies are going to be defensible. I believe they will be. The reason is the agents will codify the SOPs of the company and those SOPs, how it does business, will define the business and differentiate it in the marketplace. This differentiation will require human insight. Though we could imagine AI just trying stuff, but I’m imagining without human oversight you’d get a buggy company that is too fragile to function. We need fundamental shifts and universal models — basically AGI at that point.

You might also ask whether there is platform risk from providers like OpenAI but there isn’t. Again the value will be retained in the definition of the jobs to be done in the graph. Those jobs will define the company. Further, these jobs will store specific context and memories and get better at managing a huge variety of permutations on the base task. Effectively, you can train what to do with human intervention and correction. This will lead to agents that are ever improving. Since the workflow will be stored at the provider of the agentic workflow they will not be able to be replaced by the *LM platform company. Further since the full context is with the agentic provider it will be immensely sticky and create lock in.

At the limit it goes to the situation where the agentic provider is the trained employee base while the *LM platform is the raw untrained employment pool. The consuming end company translates the artifacts of production into value in the market. All of these layers will struggle to consume the others. The agentic provider is defending by having access to mulitple *LM platforms. The end consumer is protected by living in a dynamic world and needing to change steer the overall graph to do valuable work that can make money. The *LM platform requires massive, massive investments and huge amounts of expertise to construct and they will have data from all of the agentic providers which is an insurmountable advantage — probably. It may be possible if an agentic providers growth rises to a monopoly level they will have the data to rival the *LM providers.

It is likely this space continue to be fractured until eventually at the *LM layer and at the agentic layer monopolies emerge. At some point there may be a split at the agentic layer as it expands beyond internet usecases into the physical space or digital simulation. These spatial agentic workflows will be quite interesting. This is where true escape velocity for learning can occur because these agents can get rapid feedback in simulation. Agentic workflows for simulators and games — the right kind — could create huge opportunity for transfer learning. Language is just a proxy used to communicate shared models of the world afterall, it fails in translation because you have to calibrate these shared models against each other. This calibration happens with context but you could imagine that a world model trained in the physical world would have no need of language except to communicate with humans.

So at the limit the chaining of agentic workflows mistranslation starts to go away. Either language will be replaced by a machine native language that precisely identifies sub-sections of the neural net or language becomes ever more grounded and precise. Well, it may also be replaced by a dynamic loss calculation i.e. expectations that be searched and improved upon with gradient descent.

Overall, this is a fascinating topic. Founder-ship and management will become the end job function of humanity. Generally, you have three options as an entrepreneur, enter the market and try to win the monopoly for agentic workflow management, leverage it as a business, or build tools for agentic workflow management. Success for the second is easier as there is infinite niches, you just need to bring your domain expertise. The former is unlikely since you have to compete against startups with massive war chests of cash. The winner will need to catch on virally. The last is interesting.

Perhaps the first step is creating an agentic workflow that can create integrations since that is a thorny problem — how can we give the agents the tools to do the work? There are API hubs like Rapid API. If we could create an automated way of integrating with those APIs, construct workflows for provisioning and setting up those APIs via agentic workflow then that would provide significant value.

Next Steps

As it stands we don’t quite know enough about the challenges of agentic workflows and how to serve them to even create messaging to test the problem. To examine whether or not to get involved the next steps are to actually create tools and see what that process is like. Build a simple application and see where the problems are. In this way we can test the problem and our value proposition. If that looks good we can move on to testing the problem on the internet.