The Economist recently described a new phenomenon they label ‘agentic flooding’ – which refers to consumers using AI to overwhelm councils and corporations with appeals and refund requests. This is fueled by the rise of general and specific LLMs (there is now an AI service in the UK, TaxBandAppeal, which pulls the government’s valuation data, compares your home against every comparable property on the street and drafts a challenge letter). The rate is likened to a flood of requests, hence the term.

Unsurprisingly, in response, councils and organisations are adopting AI to parse and process these requests.

This is an agent to agent world and the key to making this work is by passing through context. Bear with me as I explain my thinking…

In the council example, there’s a shift from one to one (a person writes, a person reads), to many to many (thousands of agents writing, numerous agents reading). As with most patterns in life, we’ve experienced this before and solved it in technology.

We once coded a single connection to a single computer. Then as requests to connect mounted, the tech industry solved the overwhelm it created via the use of APIs, where the systems published application programming interfaces (essentially a defined set of functions available to other parties). The API was standardisation, and standardisation is what let many to many connections run at scale. Behind the standardisation was context. Both sides knew exactly what getAccountBalance did, what it needed as input and what came back.

As anyone who’s generated LLM output will know, the key to generating the output is the seeding of the input. That is, what documents, prompts and other context were used for the output to be generated (including the model used). As we move to an agent to agent world, receiving agents will be better equipped to respond if they have this information to hand. It will let them consider the request more deeply. Give it the seed instructions and it can do far more.

This is needed because unlike a computer system, which is defined and deterministic, the real world is undefined and extensible. We can’t write an API for the nuances of the type of correspondence most consumers and organisations engage in. So for AI to work properly, we need to find an efficient way to pass through context to the receiving party.

So I see a world where most AI-generated correspondence carries a manifest; think of it as an AI QR code or digital token attached to the letter. Inside it is everything that produced the human-readable text: the source documents, the prompt, the model, and so on.

The receiving agent reads the token before it reads the letter. Informed with the context, the back and forth shrinks, because the receiving agent already holds the context.

In effect, it’s the next step in the technology cycle: from one-to-one connections, to APIs, and now, in a natural language world, context-aware output.