Use a small model (Haiku/Gemini Flash) to continuously summarize agent events instead of traditional dashboards.
Design:
- Wrapper script (agentd-watch or similar), no changes to agent binary
- tail -f on _/events/*.jsonl, pipe to small model
- Ring buffer of last N events
- Summarize every M seconds or N new events
- Output: 1-2 line natural language summaries to stdout
Example output:
- "Agent researching, 3 tool calls, $0.02 spent"
- "Warning: 5 failed tool calls in a row"
Benefits:
- No Grafana/Prometheus setup
- Semantic understanding of what's happening
- Can highlight anomalies
- Works with existing event infrastructure
Could be as simple as a bash script with jq + curl to LLM API.