v2.1.154

Source: Anthropic Claude Code Date: 2026-05-28 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.154

Summary

A capability release anchored by two features: Opus 4.8 as the new default high-effort model (with fast mode at 2x standard rate for 2.5x the speed), and dynamic workflows — a first-party orchestration layer that can spin up tens to hundreds of background agents from a single prompt, accessible via /workflows. The lean system prompt becomes the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier. The multiple-choice question prompt is now reserved for decisions the model genuinely can’t make itself. /simplify is reworked to cleanup-only (reuse, simplification, efficiency) rather than triggering the full bug-hunting review. A cluster of background session stability fixes closes out the release.

Implications

• Fleet-as-operations-surface thread. Dynamic workflows is the first native multi-agent orchestration primitive in Claude Code — not an SDK pattern, but a built-in UX. “Tens to hundreds of agents in the background” from a single /workflows command is a step change in the surface area of what a single operator session can spawn. Fleet monitoring tooling needs to account for this new fan-out pattern.
• Model-layer holding thread. Opus 4.8 arriving with fast mode at 2x cost/2.5x speed reshapes the cost-performance curve for high-effort tasks. The lean system prompt becoming the default for new models (and not for older ones) is a quiet compatibility signal — Anthropic is managing a model generation transition in-place.
• ACP/host-slot thread. The subagent worktree-isolation guard fix (subagents in background sessions were bypassing the guard and writing to the shared checkout) is a correctness boundary that matters for multi-agent orchestration: as dynamic workflows land, the isolation guarantees between spawned agents become critical.
• Agent observability/trust thread. The auto-mode classifier improvement (better data exfiltration detection, particularly bulk repository transfers) is a trust-surface hardening that’s arriving in the same release as the tool that enables large-scale agent fan-out. Intentional pairing.