Actions, skills and modes: new design artefacts for agentic chat
We’ve recently started experimenting with chat experiences that use agentic AI. This involves working with systems that, with users’ permission, can make some decisions about how to help them meet their goals.
In AI Studio, we’re working with the GOV.UK Chat team to try out a so-called ‘ReAct’ orchestrator agent that thinks, plans, acts and reflects, adapting its behaviour as it goes. This allows it to better understand user needs and help users with more complex, multi-step scenarios, like starting a business or becoming a parent.
This shift is not only technical. It’s a design shift.
When an agent can make its own decisions, design becomes a constraint system: we shape behaviour by deciding what the agent has access to, how it should behave in different contexts, and how we evaluate whether it’s working well.
This work is in its early stages and we’re using test environments to explore it. Here’s how we’ve been approaching it so far.
Designing a ‘palette’ of behaviours
One way to think about agentic systems is as a painter’s palette. The agent can choose the colours it uses, but it can only choose from what’s available. The palette we provide influences what it can do and how it behaves.
We’re looking into different types of artefacts the agent can use – that is, the palette they have access to – both to enable actions and to make its behaviour testable.
For this work, we’re trying out the following artefacts:
- Actions - single specific actions, like ‘check MOT status’
- Skills - combinations of specialised knowledge, capabilities or processes, such as a ‘grant application’ process
- Modes - different behavioural states the system can operate in, for example ‘exploratory research mode’
These are combinations of context, instructions, tool calls or principles that the agent can adopt when needed. Each artefact includes a description of when and how it should be used.
Allowing an agent to freely choose from these raises some important design questions:
- what does ‘the right choice’ look like?
- how will choices stay legible and safe?
- how will we align with government design principles?
Each artefact is a chance to embed clarity, accessibility, inclusivity, transparency — and a way to evaluate whether the agent behaves as intended. Each artefact becomes a new design surface for us to work with, and we’ve begun to collate them in a catalogue.
Actions
For now we’re using the term ‘actions’ for single specific tasks the system can choose to take. They are discrete, executable functions that perform specific things, like ‘do a calculation’, ‘retrieve information’ or ‘send and receive data’. Each action does one thing well.
For example, for someone buying their first car, actions might include ‘Check MOT status’, ‘Register vehicle ownership’, or ‘Calculate tax cost’.
For someone becoming a carer, actions could include ‘Check carer’s allowance eligibility’ or ‘Find local support services’.
Skills
Skills contain domain expertise for certain situations. They are specialised knowledge or capabilities that guide how the system approaches complex multi-step scenarios. They shape the conversation and, potentially, orchestrate multiple actions.
For example, if someone is becoming a parent, a skill might involve asking about both parents’ employment status, explaining parental leave options, and sequencing multiple tasks (register birth → apply for benefits → get passport).
If someone is starting a business, the agent might use a skill to help someone understand the journey from hobby to registered business, combining tax processes, legal information, and practical considerations.
Modes
Modes are behavioural states the system can operate in. They adjust how the system communicates for certain situations and, potentially, how the interface is arranged.
For example, if someone is planning for their first child – considering things like nursery costs and parental leave – the system might switch to a research and exploration mode to present options, allow for creative thinking and avoid premature commitment.
Whereas, if someone is planning for retirement the system might trigger a careful compliance mode that uses extra verification, more friction, and explains statutory language.
What’s next
We plan to develop an agentic ‘playground’ – a working agentic prototype going beyond our existing Chat playground – to allow us to try out examples of actions, skills and modes, and see how they perform in a test environment.
We also want to interrogate the relationships between actions, skills and modes. For example, when are actions or skills enough, when are modes needed to shape risk and behaviour, and whether one can replace the other.