Two tools with basic schemas and minimal descriptions. start_span and end_span have input schemas with type definitions, but descriptions are generic (under 100 chars) and lack actionable context for LLM selection. Parameter descriptions are missing entirely, 'name', 'trace_id', 'parent_span_id', 'attributes', 'span_id', 'status' have no descriptions in the schema. No output schema documented. No error handling guidance. No composition hints (e.g., 'call start_span first, then end_span'). Tool names are verb-noun but lack clarity on when to use each vs. alternatives. No pagination, no batching, no idempotency guarantees stated.

• Add descriptions to all parameters in the input schema. For 'trace_id', write: 'Optional W3C trace ID (32 hex characters). If omitted, a new trace is generated. Use the same trace_id across related spans to link them in a distributed trace.'
• Expand tool descriptions to 100-150 characters with context. Example: 'Start a new OpenTelemetry span to track an agent operation. Use this at the beginning of a task, then call end_span when done. Returns span_id needed for end_span.'
• Document the output schema for both tools. Specify that start_span returns {span_id, trace_id, parent_span_id, start_time_ns, status, attributes, events} and end_span returns the same with duration_ms and end_time_ns populated.
• Add error handling to end_span. If span_id not found, return a structured error: {error: 'span_not_found', message: 'Span ID not found. Ensure start_span was called first with this trace context.', recovery: 'Call start_span to create a new span.'}
• Add a composition note to start_span description: 'After calling this, use the returned span_id with end_span to close the trace. For hierarchical tracing, pass the parent_span_id from a parent span.'