Progressive Disclosure in Agentic AI
Why doesn’t an AI agent just load every tool it has access to, every time?
That was my exact question when I first started wiring tools into an agentic project recently, and it turns out I wasn’t inventing anything new, I was rediscovering something that already had a name.
Each AI model has a limited context window, or a fixed amount of working memory. Every tool description that I give goes into the context and eats that space, whether or not that tool ever gets used.
So, what happens when an agent has access to lots of tools, say 500 tools?
Just the description of the tools, when even not using them can consume and burn lots of tokens, and agent runs out of the room before it has done anything. It feels like an AI agent stuck in the equivalent of 40 open browser tabs and a spinning fan.
Get Abhinav Gupta’s stories in your inbox
Join Medium for free to get updates from this writer.
What is the solution?
The fix is borrowed from UX design: Progressive Disclosure, which means that don’t show everything upfront and reveal detail only when needed.
When an agent starts, it just sees the skills name and description which consumes maybe < 100 tokens each. If there are 20 skills, it will be roughly around 2k combined. Full instructions loads only when a task triggers that skill.
I am watching this happen in my own tool list right now, some tools show up only as names until I search for them, and only then does the full schema loads in.
The important point is, this isn’t just saving tokens, too much context degrades output quality, not just cost. So agents pickup what they need right now, not everything it might need eventually.
But it is not free. Every on demand load is a search step, which costs latency and opens a new failure mode, for example, an agent doesn’t know that tool exists or a lookup chain nested so deep that it breaks. Indexes drift too, for example, listing skills whose underlying files got renamed or deleted.
