Validate search queries before execution
The agent receives an allow, block, reuse, defer, or advise decision for a planned web search, preventing wasted spend on invalid or redundant queries without executing the search itself.
How it works
The agent constructs a request payload containing the search query, estimated costs, and budget details. It sends this to the Agent Preflight MCP server. The server evaluates the request against rules and history, returning a decision. The agent checks the 'decision' field: if 'allow', it proceeds to execute the search via Tavily or Exa; if 'block', it inspects reason codes to fix the query; if 'reuse', it retrieves the cached result; if 'defer', it waits for budget availability. This ensures cost efficiency and query validity before any external API call is made.
Prompt
Paste this into an agent that has the servers below connected. Adjust the names and dates to your own.
Check if this search query is valid and affordable before I run it: Query: 'latest AI news', Max Results: 5, Estimated Cost: $0.01, Remaining Budget: $0.08MCP servers it needs
Source
The vendor page this playbook is drawn from, and the passage that supports it.
Agent Preflight lets an agent check a planned tool call before spending on it. It returns allow, block, reuse, defer, or advise. It never executes the destination tool.Agent Preflight MCP documentation