Who's burning your agent budget? A 6-subagent run logged 11,236 output tokens with zero attribution
qaiser_mehdi · reddit · 2026-09-03
A real-world case from Reddit: a developer's SDK usage stats only reflected the final result message — a run with six subagents was logged as input=23, output=11236, and despite logging every call, he couldn't attribute any of it to a specific step or agent.
The poster (19, researching inference cost attribution) asks teams running multi-step agents in production: can you tell which step is actually costing you money, or do you just see one number at the end and guess? "Yes we can, here's how" answers are explicitly welcomed.
More from coding & agent
- HarnessEvolve paper: dual-gate loop fixes three failure modes of self-evolving agents — dair_ai · 2026-09-03
- Mastra launches sandbox computer use so agents can browse, click, type and screenshot — Scobleizer · 2026-09-03
- How to Keep Long-Running Agents on Track: Redis State Beats Context Stuffing — Deepfeet-09 · 2026-09-03
- Looking for a <500M SLM to Summarize Code Snippets for Local Agents — Mrinohk · 2026-09-03
- LangSmith Roadshow heads to Dallas in two weeks with agent dev workshops — LangChain · 2026-09-03
- Agentic API adds a stateful layer in front of vLLM for open-model agent runtimes — techNmak · 2026-09-03