TLDR Tech
AI Agents Need Infrastructure, Not Just Models
The interesting thing about Perplexity's SPACE announcement is not the AI capability itself. It's the engineering underneath it.
Firecracker microVMs, disposable sandboxes, snapshotting, paused-session restoration. These are serious infrastructure primitives. Perplexity has essentially built a runtime that lets an AI agent work on a task for hours or months, pause, resume, and do all of that without exposing the host system to risk. That is a meaningful piece of engineering, and it signals something important about where agentic AI is actually heading.
Most of the conversation in financial services right now is about which model to use and what prompts to write. That misses the harder question: what does the agent actually run on, and who controls it?
For UK consumer credit specifically, this matters in two ways.
First, compliance. An agent that can edit files, execute code, and run autonomously over weeks is not a chatbot. It is a system that makes decisions and takes actions. The FCA's existing frameworks around systems and controls, auditability, and fairness were not written with this in mind. If your agent is processing applications, chasing documents, or making affordability judgements over a multi-hour session, you need to know exactly what it did and why. Sandboxed execution helps with isolation, but it does not automatically give you the audit trail regulators will expect.
Second, build versus buy. Most lenders and brokers will not build their own agentic runtime. They will consume one from a platform provider. That means the infrastructure choices Perplexity and others are making now will become the defaults that regulated firms inherit. Getting involved in those conversations early, and understanding what you are actually running on, is the job of technology leaders in financial services right now.
The model is almost a commodity at this point. The runtime is where the real architectural decisions are being made.
- agentic
- AI agents
- AI