Claude Code 2.1.202: Workflow OTel Attributes and Dynamic Size Control Change How Operators Govern Multi-Agent Runs

The changelog for Claude Code 2.1.202 reads like a maintenance release, but buried in the fixes are two operator-level changes that materially alter how engineering teams can govern and observe multi-agent workflow runs: new OpenTelemetry attributes scoped to workflows, and an operator-facing control for dynamic workflow sizing.

What happened

Claude Code 2.1.202 ships three notable changes for teams running dynamic workflows at scale:

1. Dynamic workflow size setting — a new /config option lets operators set an advisory size guideline (small / medium / large agent counts) that Claude applies when deciding how many sub-agents to spawn. The release notes are explicit: this is an advisory, not an enforced cap. But it is the first time this choice is surfaced as a config rather than buried in system prompts.

2. workflow.run_id and workflow.name OTel attributes — all telemetry emitted by workflow-spawned agents now carries these two new fields. Operators running OpenTelemetry collectors can correlate every tool call, model request, and token usage event back to a named workflow run rather than having to stitch together span trees manually.

3. mTLS rotation fix — transient handshake failures during in-place client certificate rotation are now resolved. This matters for any operator running Claude Code behind a mutual-TLS gateway or enterprise proxy with rotating credentials.

Additional reliability fixes include: /review <pr> reverted to single-pass (multi-agent review now requires /code-review <level> <pr#>), background agent session resume taking minutes in worktree-heavy repos is fixed, and the voice dictation unbounded retry loop on microphone failures is resolved.

Why it matters for AI engineering teams

Dynamic workflows — where Claude writes and executes its own orchestration scripts, fanning work across tens of sub-agents — have been the most operationally opaque part of Claude Code since their introduction in 2.1.160. Token consumption in a large workflow run is hard to predict and even harder to attribute after the fact.

The two changes in 2.1.202 address both problems from different angles:

Sizing advisory as a policy dial. Before, the number of sub-agents in a workflow was entirely model-determined. The new small/medium/large setting gives operators a documented knob to steer aggregate concurrency — important when your API spend, rate limits, or provider quota can only absorb a certain parallel load. It doesn't give you a hard agent count cap, but it does let you express a preference that survives prompt rewriting and context shifts.

OTel workflow attributes as billing attribution layer. workflow.run_id is a stable identifier that groups all spans from a single workflow execution. workflow.name maps to a human-readable label. For any team that needs to allocate API costs to projects, customers, or tickets, these two attributes transform workflow telemetry from a debugging aid into a billing attribution layer — assuming your AI gateway forwards OTel data and your cost accounting pipeline consumes it.

The mTLS fix is a smaller but important reliability signal: it means Claude Code is actively hardening its gateway integration surface rather than leaving edge-case TLS failures as known issues.

The router/operator angle

Dynamic workflow runs create a routing problem that most operators have solved poorly: model requests from sub-agents all look the same at the gateway level — the same API key, the same model, no workflow context. With workflow.run_id now propagated through OTel, a properly instrumented gateway can:

• Enforce per-workflow token budgets — stop a runaway workflow that has already consumed its allocation without killing unrelated sessions on the same key.
• Attribute costs to workflow names — break down your API spend by workflow type, not just by user or team.
• Audit sub-agent behavior — reconstruct the exact sequence of tool calls and model requests that comprised a workflow run, tied to a stable run ID.

The dynamic size advisory adds a complementary upstream control: if your gateway applies rate limiting by concurrency, instructing Claude to use "small" workflows reduces the burst factor before requests even reach the provider.

For the mTLS rotation fix: if your team runs Claude Code behind a mutual-TLS enterprise proxy with short-lived certificates, you should upgrade to 2.1.202 before your next rotation window. Prior versions could fail silently mid-turn during the handshake window.

What TheRouter users should watch or try

The OTel workflow attributes only create value if your observability pipeline captures them. Check whether your AI gateway or OTel collector is configured to forward span attributes from the Claude Code OTel exporter — if it is, workflow.run_id and workflow.name will start appearing in your traces automatically after upgrading to 2.1.202.

For the workflow size advisory: set the /config dynamic workflow size option to small or medium if you are operating on rate-limited provider tiers or have seen token spike incidents during large workflow runs. The advisory is not a guarantee, but it gives the model a default posture to work from.

If your team is evaluating multi-agent workflow cost governance, the combination of the sizing advisory and OTel attribution in 2.1.202 is the closest thing to first-party workflow observability Claude Code has shipped so far. Building your accounting layer on top of it now, before workflows become standard practice for your team, is the lower-effort path compared to retrofitting it later. TheRouter's API reference documents how to wire custom OTel collectors for gateway-level attribution.