Agentic Observability

Instrument every agent run so failures stop being mysteries.

Steps

1. Map the graph — orchestrator, subagents, A2A servers, MCP tools: every hop is a span.
2. Instrument — wrap agent.invoke() and every tool call with span records (timestamps, tokens, cost, tool, latency, outcome).
3. Propagate — one trace_id across the whole mission.
4. Analyze — cost per agent/tool, loop detection, failure clustering, latency percentiles.
5. Gate — eval sampled runs in CI; alert on budget/error-rate regressions.

Span record

{"trace_id":"tr_9f2c","span_id":"sp_ab12","agent":"pm","model":"qwen3:8b","tool":"web_search","duration_ms":2140,"input_tokens":812,"output_tokens":402,"cost_usd":0.0004,"outcome":"ok"}