Skip to main content

The agents are coming

How hierarchical AI agents will reshape the workplace

As we adjust to the stream of invention, innovation, and inspiration that the generative AI sector unleashes upon us daily, a recurring question is emerging: Who will be the operators of these new technologies? Who will craft the increasingly complex inputs that will be transformed by these intriguing new artificial minds into the words and images, the sounds and videos, the calculations and code, that we seek?

The need for skilled AI operators

Enter the prompt engineer. Hero of the moment. Skilled whisperer of constructed semi-prose that our new artificial workhorses so love to consume, and in return generate such impressive results.

The chances are, even the moderately tech-literate amongst us have by now dabbled with prompt engineering. That is to say we, have experimented with generative AI services in the hope of learning a little about the available capabilities and the value that they may offer in different scenarios. But alas, for the most part, amateur efforts at prompt engineering still tend to yield results that are not sufficiently high quality to represent a viable alternative to professional outputs generated by skilled humans.

To harness AI effectively, organisations will need to train, hire, or otherwise source specialist AI operators, and they face a market-wide shortage of experts with the required skills. But all this is just the start: things are about to get exponentially more complicated.

The shape of applied AI

As (some of) the dust starts to settle following the explosive arrival of generative AI, we can make several early observations and begin to look ahead at the possible shape of things to come.

First – some AI models are better at certain tasks than others, often deliberately so. Organisations that intend to comprehensively adopt AI can therefore expect to need to master multiple foundation AI models, platforms and ecosystems.

Second – organisations can also expect to operate multiple instances of any given AI base. This principle enables uniquely purposed models to be trained using increasingly specialised (often proprietary, or even personal) data sets. Some specialised training data sets will be expensive to access, and will push up the cost of models that incorporate them.

Figure 1 Simplified illustration showing how specialised models and agents might be composed to fulfil specific organisational tasks

To harness AI effectively, AI operators need technical familiarity with the platform underpinning each model and agent, and familiarity with the professional domains where specialised models and agents AI are configured.

Figure 2 Example of specialised AI models / agents in a typical organisation

Now organisations face the triple challenge of:

  1. Growing urgency to adopt AI in order to remain competitive
  2. Growing complexity in the required skillsets, and
  3. Growing shortage of the skilled operators needed to make it happen.

Things are starting to look difficult.

But wait! Aren’t AI input commands comprised of the same creatively arranged words that we have just taught AI to produce? What if we could use AI to create and fine tune the instructions that other AI models consume? Could we then outsource the task of AI operation to…AI? Well, in short, yes.

The automation of AI agent management

Alongside services that employ AI for all manner of digital content generation, the race is already on to develop the next generation of AI services that move us beyond single task execution to the AI controlled management of other AI services.

A structure like this might involve a single managing AI agent being assigned a broad task of achieving some objective, for which it is optimised to first generate an overall plan that requires skilled input from several parties, and then pass instructions to subordinate agents (or even humans) who are each optimised to handle discreet parts of a broader task (e.g. code generation, video production, legal analysis etc). These managing agents should be capable of receiving, evaluating and responding to subordinate outputs and feedback, and issuing further instructions in response.

There is some way to go before such hierarchical AI networks are viable for mainstream use. Some of the hurdles that engineers are currently grappling with include:

  1. Workflow modelling and command sequencing that enable AI agents to interpret broad instructions and generate a series of tasks to be executed sequentially by specialist AI agents without the need for interim human feedback at each step.
  2. AI input interfaces and prompt generation which enable easier, more intuitive mechanisms for interacting with AI models and agents.
  3. Agent to agent integration, interaction management, task assignment and monitoring.
  4. Agent output quality review, instruction revision and task iteration as required, and finalisation (one of the hardest elements).
  5. Output compilation that combines outputs from multiple agents / activities into a final work product which can be further scrutinised.

Evolving AI from individual request-response engines to complex multi-step activity management engines is an ambitious goal, but it’s far from unrealistic. If it works, AI slides quietly into the management domain.

Figure 3 Example of specialised AI models / agents in the management domain

This is the point where AI moves from ‘useful’ to ‘mind-blowing’.

To illustrate this further, let’s look at what an activity sequence might look like for a hypothetical business scenario across the three maturity phases illustrated above. The modelled activity is based loosely on the hypothetical execution of a 3-step action plan involving image and video generation, legal consultation, and financial planning.

Figure 4 Changing distribution of work from higher cost to lower cost of execution

One clear benefit of increasing agent adoption is a transition of activity from higher cost to lower cost execution domains. The potential scale and impact of this transformation over the next two decades is hard to overstate.

So where does this end?

Good question! There is certainly no reason to assume that advances in AI will stop once basic management capabilities have been mastered.

As we look further ahead, there is even less certainty and even more speculation, but here are some of the areas our R&D teams are busy exploring:

  • AI all the way up the value chain – are we ready for an Artificial CEO?
  • AI agents managing human team members. For years the ‘wisdom’, backed up by many studies, has been that humans prefer interacting with humans over robots. Now that is changing, fast.
  • Autonomous Economic Agents (AEA). If the idea of AI agents managing other AI agents / humans blows your mind, just wait until the AEAs arrive. Combining the self-controlling features of management agents with the independent economic identity that digital wallets and decentralised financial networks can provide. AEAs can send and receive digital assets, transact and trade with others, and accumulate (or spend) wealth.
  • How will society prepare for the longer-term socio-economic impact of hyper-automation?

Final thoughts

Hierarchical agent models have the potential to unlock massive cost and time saving benefits, but the evolution to AI-enabled business is a significant transformation process.

In the near term, the reduction in production cost achievable through targeted AI and AI agent adoption looks set to generate a windfall to early adopters in the form of improved margin. However, as AI becomes widely adopted, this early margin opportunity may give way to competition and price discounting, at which point laggards may face existential risk.

Also - whilst this blog has focussed mainly on tech and business operations, non-technical areas such as legal (IP, liability), ethics, governance, policy and regulation, data privacy, confidentiality and security all contain significant challenges that will influence the pace and direction of growth in this space.

If you are interested in learning more about any of the topics discussed in this blog, including prompt engineering, specialised AI models, and hierarchical AI agents, please do get in touch.

Further reading / some examples




This list of projects