The Post-AGI Era

AGI has arrived. Or more precisely, machines can now reason across domains, adapt to unfamiliar problems, and perform complex tasks with a meaningful degree of autonomy.

This does not mean that AI has become perfect, universally superhuman, or that intelligence has reached some final state. AGI is better understood as a threshold of generality and autonomy. What makes crossing it significant is not simply that AI has become more capable, but that it allows us to ask a fundamentally different question about how work and organizations themselves should be designed.

The immediate counterargument to such a statement is that if AGI has arrived, where are the enormous economic and societal changes that were supposed to accompany it? The answer is that model capability and economic adoption are moving on very different clocks. AI has already begun to materially affect productivity, work, capital expenditure, and entire industries around semiconductors and data centers, but most organizations still use these systems through what is fundamentally a pre-AGI framework, with AI acting as a tool that makes humans better at performing existing work.

Part of the confusion also comes from the meaning of AGI itself. Historically, the term referred broadly to intelligence capable of operating across domains rather than being restricted to a narrow task. As models have approached and crossed many of those earlier thresholds, however, the term has gradually accumulated additional expectations such as near-perfect reliability, universal superiority, or something approaching artificial superintelligence. The goalpost has moved alongside the technology.

We are also still extraordinarily early. The model I consider to have crossed this threshold, Astra, was released only recently as of this writing, while even more capable systems are already being developed. At the same time, AI has already crossed what I would call the “good enough” threshold for many people, the point at which it can meaningfully assist with most of the computer-based work they encounter day to day. Once that happens, further increases in intelligence can feel incremental to the user even when the underlying capabilities are changing dramatically.

Good enough AI and AGI are nevertheless fundamentally different. Whereas good enough AI allows a human to perform existing work more efficiently, AGI makes it possible to ask whether the human needs to remain inside that workflow at all. That distinction is where the post-AGI era begins.

That is why with the emergence of AGI-level models, a new form of company becomes possible. Rather than adapting AI to existing human workflows, these organizations can begin from the opposite direction by decomposing high-skilled work into systems of autonomous agents that collaborate, evaluate one another, learn from their performance, and continuously improve the environment in which they operate. In effect, the organization itself can enter a recursively self-improving loop.

We are therefore in an unusual moment in history. AGI-level systems are now attainable while much of the economy being built around them still follows a fundamentally pre-AGI model, creating tools for humans rather than organizations designed around autonomous intelligence itself.

To give an example of how we approach this at Infinite Ascent, we started by building a harness and trading agents to operate in the markets, treating these as the first roles within the organization that we could automate. The next step was to create reviewers of the agents’ performance, which led us to develop a meta-harness that evaluates their work and reinforces the most effective behaviors of the best-performing agents. At the same time, we deliberately attempt to preserve more creative agents even when their performance over a particular period is weaker, so that the evolutionary process does not simply converge toward a local maximum at the expense of exploration.

Moving higher through the hierarchy, we believed that the agents should eventually be able to improve the code and infrastructure on which they themselves operate. We therefore implemented what we regard as an IT department that audits the systems used by both our trading and reviewer agents, identifies problems, and improves the infrastructure supporting their work. The next step is to extend this process further down the stack, using the performance data generated by the organization to post-train open models and reinforce the behaviors that prove most effective.

The more consequential effect of AGI will therefore not be felt when AI becomes merely a better assistant to humans, but when increasingly autonomous systems begin operating substantial parts of organizations themselves, learning from the work they perform and using those learnings to improve the systems that enable their work.

An interesting asymmetry is emerging between the capabilities of frontier AI systems and the organizations built around them. The research companies developing the most capable models have largely continued to commercialize them through tools and services designed to make existing human workflows more efficient. In that sense, much of the business infrastructure surrounding post-AGI intelligence still follows a fundamentally pre-AGI model.

There is also a structural reason why this may persist. The frontier labs benefit from remaining horizontal providers of intelligence across many industries. Moving deeply into the operating layer of individual verticals would increasingly place them in competition with the very companies that depend on their models. Model development itself does not create the same conflict, which may help explain why recursive improvement has so far been concentrated most heavily at the model layer rather than in autonomous organizations built on top of it.

The same is true across much of the broader economy. Many organizations already possess enormous amounts of historical workflow data, institutional knowledge, and continuously generated performance data that could provide the foundations for increasingly autonomous systems. Yet most have so far approached AI primarily as an efficiency tool rather than as an opportunity to redesign the organization itself.

This creates an opening for new companies that can be built around post-AGI architectures from the ground up. They do not need to retrofit autonomous systems onto decades of inherited workflows, organizational structures, and incentives, and can instead design the company around artificial intelligence from the outset.

This window of opportunity will not stay open for long, as these architectures prove their value and established organizations begin adopting them more aggressively. Until then, however, the gap between what the technology can enable and how organizations are actually using it represents a significant moment of arbitrage.

The first-mover advantage lies not only in building the architecture earlier, but in the data, experience, and iterations that accumulate as the system continuously improves itself. The organizations that begin this process first may therefore be able to outpace later adopters even after the underlying approach to creating post-AGI companies becomes widely understood.

The most capable AGI-level models today remain closed, with OpenAI and Anthropic currently operating at the frontier with even more capable models available to them internally. Open-weight models still lag the closed frontier models on the most demanding cross-domain tasks on par with Astra and Fable, but if current trend lines hold, I expect open models competitive with the closed frontier to begin emerging within the coming months.

More capable open models will allow post-AGI organizations to extend RSI deeper into their own intelligence stack, with far greater control over how their systems are deployed, specialized, and improved.

For one, reliance on research labs and closed-source models creates a significant concentration risk. As these models become embedded across every layer of the organization, the labs that control them gain an outsized degree of influence over how the company operates, leaving us increasingly dependent on decisions, access, and capabilities that remain outside our control.

This poses a significant challenge for any organization operating in critical or high-risk environments, from cybersecurity and finance to defense. In these domains, organizations cannot afford to depend on model developers’ shifting judgments about what uses are permissible, safe, or aligned with their values in order to retain access to the most capable models. Critical capabilities cannot ultimately rest on policies, restrictions, or strategic decisions made by an external organization.

Second, reliance on closed-source models creates a fundamental information asymmetry. You are exposing some of your organization’s most valuable workflows, research, and intellectual property to systems controlled by another company, one that is simultaneously expanding the capabilities of the infrastructure on which your own product depends.

This does not require malice. Model companies inevitably learn from how their products are used, where their systems fail, which capabilities users demand, and which workflows become economically valuable. Over time, the infrastructure on which your product depends can itself move upward into the application layer and begin competing with what you have built on top of it.

There is also a deeper epistemic problem of provenance in frontier models. Neither users nor, in many cases, the labs themselves can fully trace where every capability, insight, or piece of knowledge originated. The Navier-Stokes controversy illustrates how difficult it can become to determine whether a model arrived at an insight independently or whether its reasoning was influenced by material it had previously encountered. For organizations working in strategically sensitive domains, that uncertainty alone is significant.

Third, and perhaps most importantly, closed models limit how deeply an organization can extend its own recursively self-improving loop. If an organization wants to continuously improve based on its own performance and data, it needs to be able to push those learnings back down through the entire AI stack.

With closed models, that feedback loop can improve the application, tools, context, memory, orchestration, and harness surrounding the model, but the organization does not have full control over the underlying intelligence itself. The most valuable performance data it generates therefore cannot be freely converted into changes at every layer of the system.

True self-improvement therefore requires control of the model layer itself, including the ability to post-train, specialize, and continuously reshape the intelligence based on what the organization learns rather than treating the model as a fixed endpoint.

Just as two people or organizations can begin with similar capabilities and diverge through experience, two agents can begin from the exact same AGI-level base model and become increasingly different systems. Agent A can be optimized through its own performance and data for one domain, while Agent B can be optimized for another. Both remain generally intelligent, but each can continuously deepen its capabilities through what it learns from performing its particular task.

It is important here to dispel one of the false assumptions often associated with AGI, or even artificial superintelligence (ASI): that sufficiently intelligent systems will, by definition, be near-perfect at any task from the outset.

That assumption implicitly treats intelligence as having a final peak. History, however, suggests otherwise. At the end of the nineteenth century, some physicists believed the fundamental structure of physics was largely understood, only for relativity and quantum mechanics to reveal