Why Claude Code’s ‘Advisory’ Setting Isn’t a Hard Cap – Uncovering the Product Logic

1. What the v2.1.202 update actually changes

v2.1.202 was released on 2026‑07‑06. The version adds only two items that affect dynamic workflows; the other 15 changes are bug fixes.

Change 1: a new /config entry called Dynamic workflow size with three options (small / medium / large). The release notes label it as an advisory guideline, not an enforced cap . Selecting an option does not impose a hard quota; it merely hints Claude about the preferred scale.

Change 2: two new OpenTelemetry attributes for workflow‑spawned agents: workflow.run_id: a unique identifier for a workflow run. workflow.name: a human‑readable name for the run.

These enable reconstruction of a run’s activity from telemetry data.

Although the changes look minor, they fit into a broader product narrative.

2. The five key milestones of dynamic workflows (six versions)

Dynamic workflows were introduced in v2.1.154, promising "tens to hundreds of agents" orchestrated automatically. Subsequent versions address mis‑triggering, scalability, and observability:

v2.1.154 – Capability launch : agents are created automatically; the /workflows view shows their status.

v2.1.157 / 158 – Mis‑trigger mitigation : a /config switch disables automatic triggering when the word workflow appears; the keyword is later renamed to ultracode to avoid accidental launches.

v2.1.186 / 198 – Scaling infrastructure : the agent list gains status filtering (running/idle/failed/done) and sub‑agents run in the background by default, making large‑scale runs feasible.

v2.1.202 – Controllability addition : the advisory size setting and the two OTel attributes together form a soft‑control solution.

These milestones can be summarised as:

Capability rollout (v2.1.154)

Mis‑trigger governance (v2.1.157 / 158)

Scaling foundations (v2.1.186 / 198)

Soft controllability (v2.1.202)

3. Advisory vs. enforced caps

The phrase "advisory guideline, not an enforced cap" carries three implications:

User experience: it is not a quota that throws an error when exceeded.

Technical constraint: the wording uses guideline instead of maximum or limit, signalling a preference rather than a system boundary.

Claude’s perspective: the setting influences Claude’s internal orchestration logic, not a guarantee to the user.

Claude may still exceed the hinted size if task complexity demands it, because the core value proposition of dynamic workflows is Claude’s autonomous scaling.

Comparing the two approaches:

User experience : enforced caps block execution with an error; advisory settings merely express a preference.

Flexibility : enforced caps are rigid; advisory settings allow Claude to deviate.

Failure mode : enforced caps produce hard errors; advisory settings may silently diverge from the hint.

Suitable scenarios : enforced caps suit strict cost/resource control; advisory settings suit psychological expectations without breaking dynamic behaviour.

Alignment with dynamic workflows : an enforced cap conflicts with Claude’s self‑orchestration, whereas an advisory guideline is compatible.

Anthropic likely avoided an enforced cap for three reasons (inferred from the timeline):

Maintaining the promise that Claude decides scale autonomously.

Avoiding complex failure handling when Claude’s judgment exceeds a hard limit.

Recognising that most users prefer a soft expectation rather than a strict ceiling.

4. The OTel attributes – handing observability to the user

The two new attributes, workflow.run_id and workflow.name, solve a concrete problem: without them, telemetry streams contain isolated agent events with no way to group them by run.

With run_id, downstream systems such as Jaeger, Honeycomb, Datadog, or Grafana Tempo can aggregate events per run, reconstruct the full activity chain, count spawned agents, measure total latency, and pinpoint failures. workflow.name provides a readable label for UI searches.

This observability investment is a classic developer‑first design: Anthropic supplies raw data and leaves governance (alert thresholds, cost attribution, trend analysis) to the user’s own stack.

The combination of the advisory size hint and the OTel attributes forms a complete controllability solution:

Advisory setting – expresses a preferred scale (no guarantee).

OTel attributes – enable precise measurement of the actual scale in the user’s observability pipeline.

Consequences of this design:

High entry barrier: not all developers have an OpenTelemetry backend.

Post‑run rather than pre‑run control: you can detect overspending only after the run completes.

Governance responsibility is outsourced to the user, matching Anthropic’s developer‑first philosophy.

5. What the whole timeline tells us

From v2.1.154 to v2.1.202, Anthropic made at least four product decisions for dynamic workflows:

Enable the capability (154).

Prevent chaotic auto‑triggers (157 / 158).

Lay the infrastructure for scaling (186 / 198).

Provide a soft controllability layer (202).

This progression reflects a broader industry shift: AI programming tools are moving from single‑agent capability races (2024‑2025) to a phase where scalability and controllability are equally important.

Dynamic workflows represent a qualitative leap—from a single agent writing code to Claude orchestrating dozens or hundreds of agents. The new challenges—uncontrolled scale, unpredictable cost, and opaque execution—are addressed by giving users a preference knob and the raw telemetry needed to build their own safeguards.

Whether this approach will persist depends on market pressure; enterprise customers may eventually demand hard caps or built‑in cost guards. For now, Anthropic’s philosophy is clear: maximise capability, minimise built‑in control, and let developers implement the governance they need.

For individual developers, the practical takeaway is to ask two questions before using dynamic workflows: does the task truly require dozens of agents, and do you have a way to stop runaway costs? The advisory setting can only hint at scale; the OTel attributes are useful only if you have an observability stack to act on them.

In short, Claude Code gives you the steering wheel (advisory knob) and the dashboard (OTel data), but you still have to drive.

Thanks for reading – feel free to like and share!