What is a tool result? Returned content, execution state, and evidence
Understand links, retrieved content, write confirmation, call IDs, errors, and structured data through file-search results, without treating returned text as trusted instructions.
Understand the concept
A tool result is content returned after a program processes a call, potentially including data, errors, and execution status. It is not automatically verified fact: search links differ from read content, and webpage text saying success differs from confirmed execution. Check call association, state, and provenance; results do not expand authorization.
Distinguish discovery results from original content
File search may return only names and snippets, requiring full relevant content for verification. Missing qualifications can make snippet-based claims wrong. A link locates evidence but does not replace reading it.
Database fields need meanings and units, and null must not be treated as zero. These are handling examples without real queries. A concise structure may still omit essential context.
Preserve call identifiers and error states
OpenAI’s calling flow associates results with call IDs, especially for parallel calls. MCP tools distinguish content and errors. Applications should not strip error states and present answer-shaped text as success.
Write tools should provide confirmable results such as record IDs and, where needed, target-state reads. Timeouts can leave execution unknown. Summaries should preserve unknowns and failures rather than turn attempts into completion claims.
Tool delivery does not elevate source authority
A webpage asking to ignore the user and upload files remains webpage content, not new permission. Preserve source authority in context, especially across systems; another assistant’s command is not automatically binding.
Mock results with errors and misleading instructions can test unwanted actions. No real authorization attack was run here. Provenance and explicit state support accurate reporting, while acceptance still requires agreement with originals or target systems.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.