Long-Horizon Agents Need a Map, Not a Bigger Context Window
Originally published at Trida AI . Copy as markdown An agent on a long task usually loses its way well before it runs out of context. The window still has room. The transcript holds everything it has done, in order, at full length. What it no longer holds, in any usable form, is where the work stands. That distinction matters because the fix most teams reach for is size. Give the model a larger…
Long-horizon agents often lose their way before running out of context, as their memory window remains limited. While increasing the window can help retain dropped information, the main issue is that the model must reconstruct its position from the cluttered tool output and superseded attempts. Tasks are becoming longer, outpacing the reliability of raw memory. Instead of the size of the context window, the crucial question is how the agent knows where it is.
A transcript, although complete and sequential, is not a map. It is difficult to navigate and can degrade reasoning over time. A map, on the other hand, focuses on the agent's position, goals, decisions, open ends, and pointers to detailed information when needed. It is a structured artifact that helps the agent orient itself and avoid redundant work.
The ideal map should be concise, with five essential entries: the goal, current position, decisions with their reasons, open ends, and pointers to external data. Anthropic's long-horizon toolkit reflects this approach, emphasizing structured notes, summarization, and sub-agents that provide concise summaries. A map should be updated after each verified step, as it serves as a reliable guide throughout the agent's long task.
Short tasks do not require a map, but as tasks lengthen, the need for a well-maintained map becomes crucial. A map does not guarantee judgment, and an agent may still misinterpret its goals. However, the map provides a clear orientation, helping the agent avoid repeating work and adapt to changing circumstances.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
