A lot of AI use revolves around agents. Unlike a simple chat completion, an agent will analyse your message, create a plan, and apply tools to achieve a task.
LLMs are stateless, so an agent needs to be able to manage context and task delegation. The common approach used to involve optimising model weights via fine-tuning or adaptors to improve task-specific performance. Now, most task-specific approaches focus on the orchestration layer, that being the agent.
To see adoption, an agent provided for a vertical application should ideally have a unique edge. But as agent workflows mature, users are beginning to expect a common set of features, to offset some of the quirks of LLMs.
From a UX perspective, this may include the ability to operate multiple agents sessions at once, all running concurrently in the background, to achieve multiple tasks at once. In order to support an agent heavy workflow, utility features like the ability to fork, rewind or similarly manage agent context. Then, also, the ability to review agent suggestions, edit them easily, and merge them logically using a version control system. Once a core agent workflow is established, differentiation for each vertical could include how a persistent memory system is managed. For example, you might want to persist common entities between multiple legal documents. It might also include modifications to or replacement of the primary workspace itself. In the case of Cursor, moving from being a VScode plugin to a native desktop application. A CLI interface is perfect for hobby users, but users with decades of experience will usually prefer a dedicated workspace such as MS Word. In a codebase or document where you need to audit every line, it's easier to do so in a workspace than in the terminal.
The combination of new agent workflows with traditional workspace tools leaves us with amusing hybrids, such as multi-agent, multi-document add-ins that sit within MS Word, combining higher levels of abstraction with familiar tools to produce a workable end product. Nevertheless, making a multi-document, browser-based, multi-agent system work inside Word is a non-trivial challenge, particularly one that is interoperable with a web app.
It is increasingly common to glue together CRUD apps using a MCP or similar connector. Custom connectors between systems that were not built bottom-up to operate seamlessly will inevitably lead to problems such as desynchronisation. There are obvious advantages to software consolidation that are compounded by the high volume nature of LLMs. In the legal technology space, we see this with underlying infrastructure tools acquiring document editors (IE Filevine acquiring Pincites or Relativity acquiring Gavel). Software companies, in an existential struggle to survive, are no longer content to own a single slice of the stack- they want to own the entire user workspace. And that is why each tool is rebuilding its own agent workflows and logic.