Anthropic’s Claude models have struggled with long‑horizon projects because a single context window cannot retain every detail. Earlier work introduced subagents, skills and agent teams, but the orchestrating Claude still accumulates all outputs, eventually hitting the same window limits. Those limits manifest as three failure modes: agentic laziness, self‑preferential bias, and goal drift. When the orchestrator stalls, the entire pipeline collapses.

Jarred Sumner of the Bun project demonstrated a practical breakthrough by converting 750,000 lines of Zig to Rust in just 11 days. He used what he calls a dynamic workflow, a JavaScript harness that externalizes the plan, lets fresh‑context Claude instances handle isolated steps, and aggregates results only at the end. The approach avoided the three failure modes and achieved 99.8% test pass rate.

The article maps six composition patterns that cover most real‑world workflow problems and warns against two costly mistakes: overusing dynamic workflows on token‑heavy tasks and neglecting proper harness design. By moving the plan into code, developers can keep each Claude focused, eliminate drift, and reliably scale medium‑to‑large projects without inflating token consumption. This pattern is already being adopted in several enterprise AI pipelines.