Custom Agents as Apps

Antigravity recently released custom agents. They are language models that can be easily re-purposed to reliably do a repeated tasks via easy-to-write harness. Let me explain.

Custom agents are agents

• Language models are the “brains” of AI apps. Custom agents are based on that exact same brain with different set of equipments.
• Repurposed to reliably do a class of tasks: The general ChatGPT app is starting to be able to actually do things for you, such as creating slides, replying to your emails, or updating a calendar events. The commonalities of these actions is that they are API calls to underlying services (google slides, gmail, or google calendar.) But that also means it requires much effort in the user to precisely let ChatGPT know what they want to achieve. This is the drawback of generality. But once, e.g., something got repeated, you can create a new harness for the language model that creates a new agent that is highly specialized to that repeated task. Due to that specialization, it has much less chance of misunderstanding the user intent and already knows the APIs needed for the tasks before hand.
• Easy-to-write harness: The two things required to create the custom agents are a prompt and a set of “skills.” The prompt tells the custom agent what its specialization is and what its skills are. The skills are the lower-level, highly open-ended types of actions the agent can take to accomplish the tasks.
  • A skill (a growingly influential concept in agentic development) is, again, a lower-level prompt instructing the language model on how to accomplishes the lowest-level tasks that constitutes the types of actions. For example, “using Gmail” is a type of actions, containing various read/write actions. Those low-level tasks are essentially known API calls, which are documented in the prompt so clearly that the LM can make the API calls itself to get the low-level tasks done.

From the original agent’s perspective, custom agents allow end-users to create “specialized” agents.

Custom agents are apps

At the same time, custom agents are actually highly-usable apps in and of themselves, taking the shape of the very same ADE software it was built out of.

The most important fact about these custom-agent apps is that they now usually cost 0 dollars to build. And this will do a very serious damage to the SaaS industry, which has been relying on the fact that people are willing to pay a positive amount of money to use those apps. And this seems to no longer be true for many apps.

The catch of this plumetting in building cost is that the apps are now much more expensive to run. In the past, an booking app cost the users and the software distributor only the infra cost to maintain the servers and databases for handling API calls. Now, the same booking app written in the form of a custom agent, beside costing those same resources, also costs inference resources (energy and chips.)

Future outlook

In the same week, DeepSeek released DeepSeek Harness, a feature allowing code generated by the agent to affect the appearance of the ADE itself, not just the texting/execution behaviors like custom agents. This feels like the next step of custom agents, where the UI can be more flexible and visually pleasing than just text.

In the coming months and years, I expect to see AI apps to evolve in this direction: making it easier for users to create custom agents AND also easier for users to use those custom agents as apps on their devices.