tool-call-flow

The tool calling cycle follows a declare-request-execute-continue structure, ensuring the runtime appends ID-paired tool-call and tool-result messages to the history before the next model turn. This protocol maintains deterministic loop control and auditability across vendor encodings.

Tool calling errors should be encoded as tool-result messages within the message history, allowing the model to read the failure and recover. This preserves the protocol-level cycle and keeps model intent distinguishable from runtime execution.

Yes, the tool-call protocol applies across Anthropic tool-use, OpenAI function-calling, MCP, and Gemini function calling. It preserves the same underlying declare-request-execute-continue cycle and ID-paired message history contract across all vendor encodings.

Parallel tool calls execute multiple requests within the same cycle, but require strict dependency limits and correct ID-paired tool-result messages. The message history serves as the source of truth to maintain correctness across concurrent streaming and execution.

Function calling breaks when the message history lacks proper ID-paired tool-result messages appended before the next model turn. Maintaining the four-phase declare-request-execute-continue cycle ensures the runtime correctly tracks model intent and execution state.

Strict message history contracts for MCP tool use are needed when implementing parallel calls, streaming, or error recovery. Defining the protocol-level cycle ensures composability and recoverability across diverse vendor encodings and bounded loop control.