Context window
The maximum text a model attends to in one call — instructions, history, retrieved material and its own output. Managing what earns a place in that window is a core engineering discipline of agent systems.
Everything competes for the same space
Instructions, tool schemas, memory, retrieved documents, conversation history and the model’s own output all draw from one budget. Deciding what earns a place is a core engineering discipline of agent systems, and the default of "include everything available" is a decision too — usually a bad one.
Bigger is not automatically better
Attention quality is not uniform across a very long context; material buried in the middle is reliably weaker than material at either end. A large window is best treated as headroom for the occasional big input, not as licence to stop curating.
Managing it deliberately
Load skills on demand rather than permanently. Retrieve memory by relevance rather than recency alone. Summarise long histories instead of carrying them verbatim. Put the most stable material first so caching can work. Each of these is a small habit; together they decide whether the tenth turn is as sharp as the first.
In AGNT
AGNT manages context per step: skills load on demand and memory retrieves selectively, keeping windows lean.
Context window — common questions
What happens when the context window overflows?
Either the call fails or something gets silently dropped — which is worse, because the agent then reasons over a truncated picture without knowing it. Trim deliberately rather than letting a limit do it for you.
How many words is a token?
Roughly three-quarters of an English word on average, so a thousand tokens is about seven hundred and fifty words. Code and unusual names cost more.
Does AGNT manage context automatically?
Yes — context is assembled per step, with skills loaded on demand and memory retrieved selectively rather than dumped in wholesale.