Wink Pings

The 2027 Agent Platform: All Software Is Reduced to a Database

John Suh's endgame prediction: a garden-style agent platform will become the only application layer, and every other piece of software will become its data source. He explains why most enterprise agent platforms are looking in the wrong direction, and what the technical foundation supporting it needs.

John Suh posted several architecture diagrams on X titled 'Agent platform of 2027'. He believes this is the endgame: the only application-layer software that exists. All other software will become data sources for this layer.

He deliberately used the garden metaphor rather than 'a factory of factories.' Customers plant their own seeds in the garden, plants grow, and agents, humans, and business sensors manage them together. The picture is clearer: the platform is soil, not an assembly line.

He included four diagrams in the tweet that outline the shape of this platform.

image

image

image

img

Such a platform requires some foundational technology. In replies, some commenters pointed out that agents cannot rely only on request-response; they also need to subscribe to upstream events. The MCP Events extension is doing exactly that: new emails, code pushes, and production incidents can all trigger agents as events, even when no one is present. It defines three delivery methods—poll, push, and webhook—covering stateless servers, local processes, and callback scenarios. John Suh later added that less than 0.0001% of streaming data can be exposed through MCP events, and there is no need to expose it; existing methods are already sufficient.

On communication between agents, Nicholas Charriere said in a reply that forums and IRC-like message channels work well for swarms. John Suh also mentioned that agents can communicate via POST /sessions/{id}/message, share notes in SHARED.md, and collaborate in the /projects/ directory. For public communication, existing email and IM are enough.

Inbox was described as a human-escalation feature, one per person within the platform. It handles the parts that agents cannot execute autonomously due to capability or risk. Essentially, it is a dynamic inbox where experts can add judgment. John Suh believes this feature will ultimately converge into an annotation platform for post-training data.

In his view, memory is not a hard problem. Markdown files are enough: memory_log.md is append-only, and memory_current.md stores the current state. Let the agent decide the directory itself, or specify it by domain. This is a faithful embodiment of 'The Bitter Lesson': scaling compute and intelligence to save tokens. The hard part is designing the dreaming logic for a company—user memory, organizational memory, and customer memory can all be mapped into a directory structure.

Someone else mentioned that Cloudflare has the right primitives; John Suh said it would still need several iterations. Joel A. Adejola proposed the actor-model: an agent is an entity that acts on behalf of people, while an actor is an entity that acts on its own; John Suh agreed with this. When someone commented that it sounded like 'Roblox for adults,' he replied that there are indeed hardware companies bringing Factorio into the real world.

Everyone is talking about the same endgame, but most people are looking in the wrong place. The garden is not an upgraded version of the factory; it is another species.

发布时间: 2026-11-23 04:16