PhD Studentship: Sovereign Agentic AI: Distilling Small Language Model Agents for Local Deployment

Agentic AI systems are language models that plan, call software tools, and complete multi-step tasks with little human input. They are moving from demonstrations into business operations such as customer support, procurement, invoice processing, and compliance checking. Most of these agents run on large models hosted by a few overseas cloud providers. For many UK businesses, this is a fragile position. Operational data leaves the organisation, cost grows with every request, and the capability depends on a supplier that the business does not control. The UK’s Sovereign AI programme, launched in 2026 with up to £500 million, and UKRI’s AI strategic framework, which names agentic AI and edge computing as areas where the UK should lead, both respond to this concern. Sovereign AI here means capability that a UK organisation owns, runs on hardware it controls, and keeps running without an external provider.

This project asks whether small language models, which run on a single workstation or an on-premises server, can become reliable agents for the bounded workflows that most businesses need. The idea is that business operations share much structure across organisations. For example, a procurement agent, an invoice reconciliation agent, or a first-line support agent works over a limited set of tools, document types, and decision rules. A general-purpose frontier model is more capable than such a task requires, and that surplus capability is what makes it expensive and hard to host locally. The project will distil agentic capability from large models into small models that keep the behaviour a task needs and drop the rest, along two complementary routes. The first route is harness optimisation.

The harness is everything around the model, including the prompts, the tool interfaces, the memory and retrieval components, the checks that verify an intermediate result, and the control flow that decides which step comes next. A well-designed harness lets a small model succeed where it would fail if asked to improvise. The student will develop methods that optimise the harness automatically for a given family of tasks under a fixed budget of local computation and latency. The second route is fine-tuning of small language models on traces of successful agent behaviour. The traces come from larger teacher models and from execution feedback that the tools return, such as whether a database query ran.

The student will study which parts of agent behaviour transfer to small models through fine-tuning, which parts are better placed in the harness, and Engineering and Physical Sciences Research Council Doctoral Landscape Award how the two should be optimised together. Throughout, the project will measure reliability, cost, latency, and data locality on realistic task suites, and will release open tooling that a UK small or medium-sized enterprise could adopt. The supervisors are Professor Shiqiang Wang (CORE-AIx Lab, IEEE Fellow, formerly IBM Research) and Dr. Tianjin Huang. The project is a collaboration with an industry research lab, whose enterprise clients face these deployment constraints. The student will join regular meetings with both supervisors and the industry collaborator, with possible visits to the industry research lab, and will publish at leading machine learning venues.

Funded by: EPSRC Engineering and Physical Sciences Research Council