Find the bottleneck executor

Background

The payoff of the trace: spotting the outlier. One executor that made 6 LLM calls while the others made 2 is the bottleneck — it's where the system spent its effort. You find it by counting llm_call events per agent_id and taking the max.

Problem statement

Implement find_bottleneck(events) returning the agent_id with the most llm_call events.

Input

• events — list of event dicts with "agent_id" and "event" keys.

Output

Returns the agent_id (string) with the highest llm_call count. Ties break alphabetically (smallest agent_id). Returns None if there are no llm_call events.

Examples

Example 1

Input:  exec_1_2 has 3 llm_call events, exec_1_0 has 1
Output: "exec_1_2"

Example 2 — no llm_call events

Input:  [{"agent_id": "orchestrator", "event": "plan"}]
Output: None

Constraints

• Count only events with event == "llm_call", grouped by agent_id.
• Return the agent with the max count; ties → smallest agent_id.
• No llm_call events → None.

Notes

• This is "debugging by measuring": instead of guessing which agent is slow, the trace names it.