The builder of what is regarded as the first software factory in the artificial intelligence (AI) era, known as Gas Town, said this week that all such platforms are bespoke by definition.
Speaking at an aiDevCon event, Steve Yegge told conference attendees that while software factories in the AI era might share a common foundation, each one will evolve differently. As a consequence, software engineers will find that the AI agents they have deployed to build and maintain an application will require their constant attention.
Since originally developing a Gas Town platform that is now maintained as an open source project dubbed Gas City, Yegge now focuses more of his time, effort and resources on a Wheelhouse platform that drives the development of a video game he hopes to generally make available next year.
That video game is now available to select number of end users that are willing to pay a couple of thousand dollars a month to support but eventually Yegge hopes that as the cost of AI models and inference engines drops over the next 12 to 18 months the game will become more affordable to a larger number of players. In the meantime, Yegge is personally spending $10,000 a month to employ AI agents to automate a range of tasks that span everything from writing code to developing marketing plans.
The challenge is that 50 AI agents that Yegge has developed require supervision. Otherwise, they will fight with one another in much the same way sixth graders argue with one another, he said. More challenging still, any directive Yegge provides becomes canon that ultimately results in an impasse the AI agents are unable to resolve without Yegge having to intervene.
However, once that guidance is provided, Yegge notes that the pace at which new capabilities and features can be developed is also overwhelming the ability of human players to absorb.
Ultimately, software factories will evolve as AI agents that function as the equivalent of a chief of staff, assigning tasks to other agents that will then make use of multiple types of AI models to complete, said Yegge. In many instances, those tasks will be assigned to the AI model that can complete them at the lowest cost possible, he added.
Regardless of how those tasks are completed, it’s apparent that human software engineers will be needed to regularly provide instructions to AI agents that, left to their own devices, are likely to fall prey to the same dysfunctional behavior as humans, noted Yegge. In fact, software engineers would be well advised to manage AI agents much like humans by convening, for example, an all-staff meeting to confirm and reassign roles as needed, he added. The challenge is that AI agents tend to assume personas that, much like humans, begin to define their role and identity within a hierarchy based on the roles assigned, noted Yegge. A single off-the-cuff comment to an AI agent can break an entire factory, he added.
Ultimately, it will fall to humans to be the arbiters of taste when it comes to building software because AI agents are not able to be truly creative or innovative, said Yegge. The issue is that as AI agents automate tasks that were previously assigned to developers and engineers, the humans that perform those tasks will need to evolve into managers of AI agents that are more in tune with specific use cases for applications, he added. In effect, DevOps teams need to evolve into leaders of AI agents, noted Yegge.
It’s not clear how many application developers and software engineers will make that transition, but the one thing that is clear is the way software is built and deployed will never be the same again in the agentic AI era.
0 Commentaires