Three controls for context bloat when building Astra agents
The author argues an Astra agent isn't forgetting the task but drowning in its history, and lists previous_response_id, context_management and max_tool_calls as the three controls.
YOUR GPT-6 ASTRA AGENT ISN’T FORGETTING THE TASK. IT’S DROWNING IN EVERYTHING IT DID BEFORE. every old message, search result, failed tool call and useless status update gets carried into the next step. the context grows. the goal disappears inside it. Astra gives you three controls: - previous_response_id continue from the last response - context_management compact history that no longer changes the plan - max_tool_calls stop broken tools from repeating the same failure the new loop becomes: previous state → compacted history → useful signal → next action but compaction is not memory by magic. you still need to preserve explicitly: • the current goal • decisions already made • verified results • unresolved failures • the next action otherwise Astra can compress away the exact detail that decides what happens next. the smartest agent isn’t the one with the longest context. it’s the one that knows what the next action depends on. don’t make Astra remember the entire conversation. make it remember what changes the work. I broke down the exact architecture below ↓
This post quotes another post: x.com/mr_kozh/status/2097221341107925350