| name | agent-loop-trace |
| description | Visualize an agent's perceive→think→act loop. Inspects tool calls, latency per step, decision branches, retry patterns. Surfaces bottlenecks. Triggers on "trace agent", "why is agent slow", "agent loop analysis", "what did the agent do". |
agent-loop-trace
Workflow
- Locate agent run logs / transcript.
- Parse into steps: each step = {perceive (input), think (reasoning), act (tool call), observe (result)}.
- Compute per-step metrics: latency, tokens in/out, tool used, error?
- Render timeline + decision tree.
- Flag patterns: retry loops, redundant tool calls, hallucinated tools, premature termination.
Output
Run: agent_xyz (total 47s, $0.18, 23 steps)
Step | t(s)| Tokens | Action
1 | 0.4 | 220/80 | perceive: user prompt
2 | 2.1 | 280/40 | think: plan 3 steps
3 | 8.2 | 120/0 | tool: search ✓
4 | 1.1 | 140/60 | think: refine
5 | 7.8 | 110/0 | tool: search ✗ (same query) ← REDUNDANT
6 | ...
Anti-patterns detected:
• Same search query twice (steps 3, 5)
• Long think gap with no progress (steps 11-14)
• Final answer not grounded in tool results
Reference
"How I use LLMs" / "Deep Dive into LLMs" Karpathy talks
