Context control in GenAI

The more I think about this, the more it resonates: whatever "Agent" means to you, and there are lots of interpretations of this buzzword, the next iteration of GenAI will likely depend on and emphasize context control.

What I mean by that is, most popular tools out there ingest and output huge amounts of content. In order to get it to achieve whatever task you have in mind, you have to herd your agents in the right direction. But that also poses a problem: how can you accurately assess their performance the further away from their output you stand?

You: Book me a holiday
Agent 1: Plan holiday
Agent 2: Assign tasks based on plan
Agent 3: Book airline tickets
Agent 4: Book hotels
Agent 5: Book museum tickets
Agent 6: Book tours
…
Agent N: All done.
You: ???

There’s a gulf between current chatbots and this kind of workflow. This open-ended question, with agents that can take action, is already possible to answer, but to get the actions you want, there is an entire tech stack still waiting to be built.

Which brings me back to context. Handing over all the context to an LLM and expecting it to "do the right thing" is naive at best and dangerous at worst. You sort of have to rein in some kind of control over it; otherwise, it runs with whatever great idea it came up with, and its author (OpenAI, etc.) assumes that it can self-correct with feedback loops and checkers. But the fundamental idea is based on probability, the most generic, most likely path, which will seldom actually align with what you wanted. It’s just given all the information above, this is what most people, most of the time would agree is the best way forward. But that’s at every turn. So you need to step in, update the context, and run again. Rinse, repeat.

Instead of trying to build this context away from humans, in some perfect database, I think that the challenge is bringing the context closer to humans, and making it easy to read it and control it, and having granular contexts at every level of the agent stack.