01
Multimodal Perception
How can agents build a coherent understanding from different kinds of information?
Potential application: Connected understanding across documents, data, and media.
Research
Towards a new generation of specialist AI.
We envision a world where specialist AI agents expand what people and organisations can achieve—bringing deeper understanding, continuity, and coordination to complex work.
Our vision
From individual tools to coordinated intelligence.
Our ambition is to move beyond AI that responds to isolated requests, towards specialist agents that understand their working context, carry knowledge forward, and collaborate across complex tasks.
We see these agents becoming a new layer of organisational capability: supporting human judgement, connecting expertise, and helping people turn ideas into sustained action.
Our research explores the foundations of that future—from perception and memory to reasoning, learning, and coordination—with human oversight and clear operating boundaries built in.
Research directions
Nine areas. One connected ambition.
Together, these nine areas explore how specialist agents can understand, reason, act, and collaborate—and become more capable and dependable over time. Each pairs a foundational research question with a potential practical application. These questions guide our research; capabilities are evaluated within each implementation.
02
Memory
How can agents carry relevant knowledge forward as their working context evolves?
Potential application: Continuity across tasks, sessions, and evolving workflows.
03
Reasoning
How can agents reason through complex situations, weigh evidence, and recognise uncertainty?
Potential application: Evidence-based decisions with explicit assumptions and uncertainty.
04
Planning
How can agents turn broad objectives into coherent plans and adapt as conditions change?
Potential application: Adaptive execution as priorities and conditions change.
05
Tool Use
How can agents use tools effectively to translate understanding into action?
Potential application: Effective integration with existing software and tools.
06
Coordination
How can specialist agents and people combine their expertise towards shared goals?
Potential application: Coordinated teamwork across people and specialist agents.
07
Learning
How can agents learn from experience and feedback while preserving dependable behaviour?
Potential application: Feedback-driven improvement across repeated work.
08
Evaluation
How can we assess agents’ capabilities, understand their limitations, and track meaningful progress?
Potential application: Measurable capability and visible limitations over time.
09
Safety
How can increasingly capable agents remain within clear boundaries and under effective human oversight?
Potential application: Controlled operation with permissions and human review.