Function call_tools 

pub async fn call_tools(
    ctx: &AgentContext,
    tool_calls: &Vector<ToolCall>,
    max_concurrency: usize,
) -> Result<Vector<Message>, AgentError>

Executes multiple tool calls and returns the results as messages.

Returns tool response messages suitable for continuing an LLM conversation, always in the same order as tool_calls regardless of completion order. Before execution, each call’s arguments are validated against the tool’s JSON Schema with lightweight type coercion (see [coerce_by_schema]); the coerced arguments are what the tool receives. A failure of one call (unparseable parameters, schema validation failure, or a tool error) does not abort the others: it is returned as a tool message with is_error: Some(true) so the LLM can recover.

Scheduling follows each tool’s ExecutionMode:

• Consecutive calls to Parallel tools are executed concurrently, with at most max_concurrency calls in flight at once.
• A call to a Sequential tool (the default, including tools not found in the registry) acts as a barrier: every earlier call must finish first, and the sequential call runs alone.

Cancellation

When ctx carries a cancellation token (see AgentContext::cancel_token) and it fires mid-execution, in-flight calls are dropped and every call without a real result receives a synthetic is_error tool message with content "Operation aborted", carrying the original tool_call id. Calls that already completed keep their real result. This keeps the message history consistent: the model sees a result for every call it issued, and finished side effects are not misreported as aborted.

Arguments

• ctx - The agent context for the invocations
• tool_calls - The tool calls to execute
• max_concurrency - Upper bound on concurrently running Parallel calls; values below 1 are treated as 1

Returns

A vector of tool response messages, one for each tool call, in input order.