agtail vs OpenTelemetry

Modern coding agents increasingly emit OpenTelemetry — Claude Code, Codex, and Gemini CLI all have built‑in support (metrics, traces, logs); others vary. So if the telemetry already exists, why read local transcripts at all? Because OTel and agtail answer different questions.

Two different jobs

• OpenTelemetry → Datadog / Grafana = observation. Aggregate metrics and structured events across your whole fleet: cost, tokens, latency, error/tool‑failure rates, adoption, trends, alerts. Great for "who is using AI, how much, at what cost and quality, trending which way."
• agtail = forensics. The full content of an individual session — every message, the model's thinking, each tool call with its inputs / results / diffs, hooks, and subagents — searchable and reconstructable, read‑only, from the files already on disk.

What OpenTelemetry does better

• Fleet dashboards, trends, and alerting.
• Localizing: drill from a cost or error spike down to the user / model / session, even the expensive turn (via session.id and per‑request cost events).

agtail is not a metrics tool — if that is your goal, use OTel.

Where agtail is the right tool

Even for a drill‑down like "which prompt blew up the cost," OTel gets you to the session and turn — but to see why (the actual tool output, the giant file Read, the loop, the accumulated context that ballooned tokens) you need the full transcript. And in practice:

• The content usually isn't in OTel. Prompt/tool‑content logging is off by default (privacy, cost, egress), so the text often isn't shipped — or is truncated/sampled.
• Retroactive. agtail reads sessions that already ran, including before any instrumentation was enabled. OTel only has what it captured going forward.
• No setup, and cross‑tool. Most CLI agents now support OTel (Claude Code, Codex, Gemini CLI all do), but it has to be turned on ahead of time with a collector and a backend — which individuals rarely do. agtail reads any tool's local transcripts with zero configuration, including tools that emit no OTel (custom scripts, some IDE integrations) and, crucially, every session that ran before you set anything up.
• Local & private. Nothing leaves the machine — no backend, retention, or per‑GB cost. OTel ships content to a backend.
• Faithful & complete. agtail reads the raw files: full outputs, diffs, thinking, subagent threads — nothing truncated.
• Zero setup. No instrumentation, collector, or exporter to configure beforehand.

Rule of thumb

• Want metrics, dashboards, and alerts across a team → OpenTelemetry + Datadog / Grafana.
• Want to read, search, and reconstruct what an individual session actually did — completely, retroactively, across tools, and locally → agtail.

They compose well: OTel spots the anomaly or trend; agtail opens the offending session and shows you what actually happened.