This post from the OpenHands blog introduces the OpenHands context condenser, which it presents as an advance in the OpenHands agent architecture aimed at managing growing conversation context. It starts from a problem familiar to anyone who has run an agent on a long, complex task: as the conversation grows the agent becomes slower, more expensive per turn, and less effective, because LLMs do worse with lots of irrelevant information in context. Starting a new chat avoids this but sacrifices continuity and leaves the user to carry context across sessions by hand.
The condenser keeps the conversation bounded instead. Once the conversation grows beyond a threshold, older interactions are summarized while recent exchanges are kept intact, giving the agent a concise “memory” of what happened earlier without retaining every detail. The post headlines three results: up to 2x lower per-turn API cost, consistent response times in long sessions, and equivalent or better performance on software engineering tasks.
Key Points
- The summary is meant to encode the user’s goals, the progress the agent has made, and what still has to be done; for software development tasks it also preserves technical details such as critical files and failing tests.
- Condensation fires only when the context reaches a specific size, which the post says keeps the benefit of prompt caching by amortizing the cost of rebuilding the cache across multiple turns.
- When events are of similar size, the post argues, baseline context management without condensation scales quadratically over time while the condensed approach scales linearly.
- Evaluated against the baseline OpenHands agent on a subset of SWE-bench Verified instances, average per-turn cost diverged from the baseline once condensation kicked in and settled at less than half of it.
- On that subset the condensing agent solved an average of 54% of instances against the baseline’s 53%, and solved a larger share using less money, fewer tokens and less completion time; the post says these trends hold regardless of the breakpoints chosen.
- The one trade-off reported is a higher number of turns, which the post attributes to the agent occasionally needing a turn to condense its context.
- All of these results come from OpenHands’ own evaluation of its own agent, reported in a post announcing the feature.
Context
The post frames the condenser as a shipped feature: context condensation is available in OpenHands, including in the managed OpenHands Cloud, where it is described as ready to use. It positions summarization-based condensation as an alternative to starting a new chat, and treats cost, latency and effectiveness as the three things long contexts degrade. For other approaches to summarizing an agent’s history as it nears its limits, see Compaction.