Vince Trost

The through-line

Vince Trost's central argument is that the industry has anthropomorphised AI agents into the wrong shape. "when we think of Agents we think of them as being these permanent persistent uh things with identity," he says, and that instinct is understandable, because "we're humans and the only other agents we really know are also humans" [1]. But the analogy fails on inspection: his own trainer has a life outside training him, whereas "AI agents don't have lives outside of work they aren't conscious beings and they don't have a really good reason to exist Beyond a task to be completed" [1]. His working definition is deliberately spare, an agent as "a system that can pursue a goal," with the goals still supplied by people [1]. Everything else in his thinking follows from stripping persistence out of the picture and rebuilding agents as disposable, assembled on demand, and thrown away when the task is done. He is blunt enough about it to title the argument "agents are trash" [1].

On memory as construction, not property

Trost starts from a technical fact that he thinks is under-absorbed: "large language models suffer from short-term memory loss they don't actually remember anything about you as you use them they're just generating the most probable response based on the context that you provide them" [1]. The consequence is that memory is never a property of the model; it is an artefact developers manufacture. "anything developers do to create some illusion of innate memory has to be constructed and that construction is all about providing the language model with the appropriate context" [1]. That context breaks into instructions on how to complete a task, examples of successful outcomes, and any additional information missing from the system [1]. He acknowledges the burden this places on builders and waves it off: "which sounds like a lot and it can be overwhelming but I actually think it's fine" [1]. There is real research on what context improves model responses, but his interest is in the practical consequences that he thinks get overlooked [1].

On just-in-time assembly

The alternative to a persistent agent is one composed at the moment of inference. Rather than "Agents sitting idle on computers waiting for instructions," resources and data can be "provisioned on demand to assemble agents with the exact right context whenever you need it" [1]. He frames this first as an efficiency argument, then as an architectural one: the approach "creates a very modular and composable world," where developers "string together unique pieces" instead of rebuilding the substrate themselves [1]. Assembling "someone's identity just in time for inference" means including only what is immediately relevant and necessary to get the desired result, and discarding it afterwards [1].

On how small agent developers win

Composability matters to Trost because of a competitive reality he states plainly. If agent developers "are going to have an edge against the big players they're going to need to pick a very specific task and they're going to need to make sure that they're better at that specific task than the best general purpose models out there" [1]. Modular infrastructure is what buys them the room to do that, letting them "focus Less on those things and more on their core service" [1]. This is the commercial logic behind what Plastic Labs builds, which he describes as infrastructure for more personalized AI agents, aimed at agent developers [1].

On privacy and the inference trail

Trost treats model inference as a privacy threat rather than a feature. Humans "will often withhold information in order to avoid revealing too much about ourselves," and he argues we should want the same discipline with language models, because research shows how powerful their inferencing capabilities are [1]. His example is deliberately mundane: mention jug handles while describing your commute and a model will "effortlessly infer that you're driving in New Jersey which is a pretty deep insight into your life" [1]. Scale that up and the danger compounds, since "thousands of other effortless but in between the lines inferences" create "a trail of data that can be very dangerous and Ed adversarially against you" [1]. Disposability is the mitigation. Context that is assembled for one inference and then thrown away leaves no trail to exploit [1].

On containment as a safety strategy

The same architecture doubles as a safety argument. Trost's claim is one of blast radius: if something goes wrong, "it's contained to one user on one specific task and that single inference can be systematically debugged" [1]. Narrow, disposable, task-bound agents fail narrowly, and the failure is inspectable at the level of a single inference rather than tangled into a persistent entity's accumulated state [1].

On what disposable agents do to the economy

His forecast borrows from market structure. "much like the Advent of trading Bots created an explosion of of trading activity on stock markets disposable AI agents will also create an explosion of economic activity in any Market that can be addressed with language and code" [1]. He declines to oversell the destination, saying he is "not sure leads to Utopia necessarily," but expects it to "create a bunch of remarkable new kind of work for people to do" and calls that "a very exciting future to live in" [1].

Takeaways

• Persistence is the wrong default for AI agents; they have no life outside the task and should be built as disposable, task-specific systems [1].
• LLMs have no innate memory, so any sense of continuity is constructed by developers through instructions, successful examples, and missing information supplied as context [1].
• Assembling an agent's identity just in time for inference improves efficiency, since resources are provisioned on demand rather than sitting idle [1].
• Modularity and composability let small teams pick one narrow task and beat general purpose models at it, which is the only viable edge against large incumbents [1].
• Models infer far more than users disclose; a mention of jug handles reveals you are driving in New Jersey, and thousands of such inferences form a trail that can be used adversarially [1].
• Discarding context after a single inference limits both privacy exposure and safety risk, since failures stay contained to one user and one task and can be debugged systematically [1].
• Expect disposable agents to trigger an explosion of economic activity in any market addressable by language and code, comparable to what trading bots did to stock markets [1].