An agentic system uses Claude with tool_choice: {type: 'auto'} to call a set of 15 tools across a multi-step research task. The system processes a complex user request and Claude calls tools in sequence, with each tool result returned in subsequent messages. After 8 tool calls, the context window approaches 190,000 tokens. The system must complete the remaining analytical steps. Which approach correctly handles near-context-window situations in production agentic systems?

Context compression via summarization is the correct production pattern for long-running agentic loops approaching context limits. The implementation: before the next tool call, inject a meta-instruction for Claude to summarize what it has learned from the accumulated tool results, replace the verbose tool call/result blocks with a compact state summary, and continue. This preserves the semantic content (what was discovered) while freeing context space for subsequent reasoning steps. The summary must retain key facts, intermediate conclusions, and task state — not just truncate. Option A is factually wrong — max_tokens controls output generation length, not the input context window limit; the 200K limit is total context (input + output combined, or model-specific). Option C misunderstands streaming — it doesn't extend context; it changes how output is delivered, not how context limits work. Option D (truncating oldest messages) is dangerous in agentic tasks — early messages often contain the original task specification, tool schemas, and critical constraints that ground all subsequent reasoning.