Agent Context Drop

Definition

Agent context drop is the loss of accumulated reasoning state when a tool call failure mid-session causes an agent to restart or diverge from its original task without retaining prior context.

In Depth

Context drop is most common with MCP-exposed search tools that return HTTP errors due to rate limiting. When an agent's search tool call returns a 429 or 503, most agent frameworks pass the error back to the LLM as a tool result. The LLM may interpret this as 'no results found' and proceed with a different approach — often contradicting earlier conclusions built on successful searches. The agent hasn't crashed; it's silently wrong. The mechanism varies by framework. In LangChain's AgentExecutor, a tool returning an exception string causes the agent to continue with that string as context, which can pollute subsequent reasoning steps. LlamaIndex's ReActAgent similarly continues after tool errors, treating error messages as observation data. Only explicit stop-on-error configurations prevent forward contamination. Rate-limited MCP search tools are the primary trigger because they fail intermittently — some calls succeed, some fail — creating partially-built context. A session that ran 8 searches successfully before hitting a rate limit has 8 results worth of grounded context followed by one failure. The agent's subsequent reasoning is built on a mixture of real and absent data. Prevention: wrap MCP tool calls in a retry layer with exponential backoff before they reach the agent. Log all tool results (success and failure) externally. Use structured error returns that the LLM prompt explicitly handles ('If you see TOOL_RATE_LIMITED, wait and retry the same query').

Example Usage

Platforms

Agent Context Drop is relevant across the following platforms, all accessible through Scavio's unified API:

• google

Related Terms