With all the recent volatility in the AI space, bottlenecks in productivity keep shifting around rapidly.
The latest bottleneck du jour that folks are converging on (now that intelligence and tool-calling have caught up) is context.
Context is what separates generic, tone-deaf, sloppy output from something magical that feels like it was made for you. Context also helps you suspend your disbelief and accept the output as good, even if it’s not perfect. “At least it gets me.”
The problem with context isn’t that people lack it. It’s that it’s scattered everywhere: in their heads, in old notes, in transient conversations, across various apps and ecosystems. Getting it consolidated and formalized into a form an agent can use is a bottleneck.
Ideally the tools and methods for capturing and routing context are lightweight, almost invisible.
One effective way I’m finding to mine context is the interview method. Instead of making myself fill out a blank document, I have an AI agent ask me questions like a podcast interview, and draw out what it needs.
This conversational approach feels more natural and less like a chore. Both agent and human can also riff off of each other and shape the direction the context-building goes.
Once the interview is done, the transcript can be used in two ways: converted into a structured summary that gets fed to future agents, or used directly as raw context. Both are valid options.
For a long time, big tech companies have been surveilling users and mining that data to gather rich insights. I think individuals are now about to do the same with themselves.
I see users collecting data and analytics about themselves as taking off. What you read, what websites you visit, how you spend your time, what you Google search — these patterns reveal things about yourself that you might not be able to consciously articulate on your own.
An agent with access to this kind of data can make much better inferences about what you actually care about, as opposed to what you say you care about. Revealed preferences instead of stated preferences.
Never before has it been more valuable to capture data from all aspects of your life.
Gathering context is one challenge. Curating it into something an agent can act on is another.
Good context curation means knowing what’s relevant for a given task and leaving out what isn’t, so that the signal doesn’t get buried. The goal is a context layer that’s curated, not comprehensive. It’s like telling a friend a quick story about what happened at the store, without telling them your whole life history.
This is still largely unsolved territory. The agents that will feel magical aren’t necessarily the ones with the largest context windows — they’re the ones that have been informed just well enough to help you meet your goals.
There’s also the question of security and privacy. What context do you want to share with an agent and a third-party intelligence provider, and what do you want to keep for yourself?
I’m excited to see the primitives that roll out in this space over the next year. Meanwhile I have some ideas to start researching and building my own.
New and featured essays sent directly to you. No spam, no fluff.