Agentplay puts coding-agent costs under microscope

agentplay is compelling observability for developers who want evidence of what autonomous coding agents actually did, but its cost telemetry is not yet reliable enough for serious FinOps.

• –Local-first architecture keeps transcripts and maps on the developer’s machine, with prompt text, tool payloads, and file contents redacted from exports
• –Replayable timelines make retries, context compactions, errors, subagents, edits, and pull requests far easier to inspect than raw JSONL logs
• –Apache-2.0 licensing and the agentplay/v1 format give other agent tools a useful foundation for interoperable session observability
• –Estimated costs rely on public list pricing and currently miss newer models, plan-specific rates, and one-hour cache pricing
• –Cursor’s deliberate omission of billed-usage lookups protects privacy but leaves cross-agent cost comparisons incomplete