{"version":1,"schema_version":"1.0.0","trace_id":"duration-fix-20260902","title":"What Happens When an Agent Works?","standfirst":"A prompt’s journey through model inference, orchestration, tool use, sandboxed execution, storage—and back again.","thesis":"An agent is a heterogeneous, iterative computing workload whose completion time depends on the full loop between model inference and environment execution.","counts":{"traces":1,"events":16,"observed_events":8,"sources":24,"claims":12,"benchmarks":6},"evidence_states":["derived","missing","observed","representative"],"limitations":["No model or paid API was called, so the trace is not a production Codex or Claude Code telemetry capture.","Model identity, tokens, inference latency, provider CPU/GPU activity, queueing, utilization, and energy are missing.","CPU time is a command total at 0.01-second source precision, not sampled utilization or loaded-core performance.","Storage figures are selected logical file sizes, not total filesystem or device traffic; network bytes were not measured.","One deterministic fixture cannot establish an industry-wide CPU/GPU workload share."],"endpoints":["traces/default","sources"]}