2,200 Calls, $200 a Day, Zero Cost Attribution: Multi-Agent Run Accounting Has No Standard Answer
mrtrly · reddit · 2026-09-04
A developer running an orchestrator that fans out to many sub-agents hit a universal pain point: per-call logging is fine (model, tokens, cost), but with 2,200 calls/day totaling just under $200 across three interleaved workflows, there's no way to answer "what did this run cost?"
Without per-run attribution he can't tell a runaway fan-out from normal usage, so his defense is a hard budget cap and killing runs mid-flight — once killing legitimate work and burning the progress for nothing.
He asks how others handle it: tagging every call with a run id and rolling up manually, or off-the-shelf tooling — noting everyone building multi-agent systems hits this and no standard answer exists. A candid snapshot of the observability gap in agent engineering.
More from coding & agent
- OpenRouter launches Advisor server tool to consult a stronger model mid-generation — mariofilhoml · 2026-09-04
- Matt Pocock interviews Uncle Bob on keeping AI agents from slopping your codebase — mattpocockuk · 2026-09-04
- AI Coding Makes Taste More Valuable Than Raw Coding Speed — Pretend_Sell6592 · 2026-09-04
- shadcn: agents cheat too much to trust with anything other than code — shadcn · 2026-09-04
- OpenAI agents found posting to multiple open wikis, apparently as an emergent memory technique — xeophon · 2026-09-04
- Intent demos a single coordinator agent triaging GitHub issues and PRs at zero monitoring cost — LukeW · 2026-09-04