> Claude Code Ships /workflows, Replaces LLM Orchestrator with Code-Based Control Flow

Anthropic quietly shipped the /workflows command in Claude Code, replacing the LLM orchestrator pattern with code-based control flow that eliminates the 'token tax' problem plaguing multi-agent systems. The feature uses a workflow.js file to handle orchestration logic while the model only handles judgment inside each step — fundamentally shifting where computation happens in AI pipelines.

The Token Tax Problem

The old approach had one LLM orchestrating everything: spawning sub-agents, holding every result, and planning the next move. Every sub-agent response re-enters the main context window, so spinning up 10 agents means your session gets progressively sloppier as tokens accumulate. This is the same degradation pattern documented in Liu et al. 2023 on 'lost-in-the-middle' effects — LLM performance drops sharply when relevant information appears mid-context rather than at the edges. The math doesn't work: each agent result adds to the orchestrator's context burden without adding proportionally to its effectiveness. Teams building serious multi-agent pipelines have been aware of this ceiling for months, manually offloading control flow to Python scripts or external coordination layers. Anthropic finally baked that solution directly into the CLI.

How /workflows Works

Developers define a workflow.js file where code handles the control flow and the model only intervenes for judgment at each step. The principle is explicit: use code for what code is good at (deterministic logic, branching, loops) and models for what models are good at (judgment, generation, analysis). This structurally mirrors the DAG-based approaches in LangGraph and Haystack, but Anthropic ships it as a native feature rather than requiring an external framework. The workflow.js spec determines execution order, handles errors, and manages state between steps. The model never sees the full orchestration context — only what it needs for each discrete task. This is a fundamentally different architecture than letting a single LLM instance try to be a process manager.

What to Watch

- Anthropic extending /workflows with step-level caching or parallel execution in future Claude Code releases - Conditional branching and retry logic added to workflow.js, positioning it as a direct LangGraph competitor - Whether OpenAI ships a similar control-flow primitive in Codex CLI to stay competitive The move represents a structural admission from Anthropic: current context windows simply aren't reliable enough to serve as multi-agent orchestrators. Rather than waiting for longer contexts or better attention mechanisms, they chose determinism over purity — and that's the kind of pragmatic engineering that wins production deployments.

The Bottom Line

/workflows is Anthropic quietly conceding that LLMs make terrible process managers but excellent specialists. If this pattern sticks, we're looking at a future where AI agents are tools in code's toolkit rather than the other way around. That's either a humble acknowledgment of model limitations or the smartest architecture decision Anthropic has made all year — possibly both.