English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Opus 4.8 and Dynamic Workflows: Claude Code Enters the "Auto-Harness" Era

Forum topic · 小凯 · 2026-05-31

Summary

On May 28, Anthropic released Claude Opus 4.8 alongside Dynamic Workflows, a research-preview feature in Claude Code that lets users describe a task in natural language and have Claude auto-generate a JavaScript orchestration script. The script can spawn tens to hundreds of parallel subagents, apply adversarial verification, and iterate until results converge. Opus 4.8 brings stronger benchmark scores (SWE-bench Verified 88.6%, OSWorld-Verified 83.4%), roughly 4x fewer unmarked code defects than 4.7, a 1M-token default context, mid-conversation system messages, tiered effort control, and a fast mode priced at $10/$50 per million tokens. The shift lowers the barrier for multi-agent workflows: routing Haiku for information gathering, Sonnet for routine analysis, and Opus for verification can cut costs 5–10x. Real-world cases include a 750,000-line Zig-to-Rust port, dead-code discovery for Klarna, and repository-wide security audits.

Key Points

  • Dual release on May 28: Anthropic shipped Claude Opus 4.8 and Dynamic Workflows (research preview in Claude Code) on the same day, signaling a shift from manual harness engineering to AI-generated orchestration.
  • Opus 4.8 reliability gains: SWE-bench Verified 88.6%, SWE-bench Pro 69.2%, OSWorld-Verified 83.4%. Anthropic reports about 4x fewer unmarked code defects compared with 4.7, making long autonomous chains more trustworthy.
  • Context and control upgrades: 1M-token default context (no beta header required), mid-conversation system messages that preserve prompt cache, and three-tier effort control (high / extra / max) without switching models.
  • Dynamic Workflows in one sentence: Describe a task in natural language and Claude writes a JavaScript orchestration script that splits the work, spawns many parallel subagents, runs adversarial verification, and iterates until results converge.
  • Three core mechanisms: parallel fan-out with isolated subagent contexts, adversarial refutation between subagents to suppress groupthink, and convergence-driven iteration that runs as many rounds as the task needs.
  • Cost leverage via model routing: Information gathering on Haiku 4.5, routine analysis on Sonnet 4.6, deep reasoning and verification on Opus 4.8. Anthropic suggests this hybrid pattern can reduce total cost by roughly 5–10x versus all-Opus execution.
  • Requirements: Claude Code v2.1.154 or newer, Enterprise / Team / Max plans (Pro has limits), and Auto Mode enabled; available on CLI, Desktop, VS Code extension, API, Bedrock, Vertex AI, and Microsoft Foundry.
  • Documented use cases: 750,000-line Zig-to-Rust port planning, Klarna dead-code discovery with adversarial validation to avoid breaking reflection or dynamic-loading paths, and repository-wide security audits on every service endpoint.
  • Triggering: Natural-language requests, bundled workflows such as /deep-research, or setting /effort ultracode so Claude decides when a workflow is needed.
  • Runtime controls: /workflows view lists active runs; pause/resume caches completed subagent state. Reopening Claude Code starts a fresh session rather than auto-resuming.
  • Pricing: Opus 4.8 Standard $5 / $25 per million input / output tokens; fast mode $10 / $50. Haiku 4.5 and Sonnet 4.6 remain available for script-level routing.
  • Competitive snapshot: Anthropic automates orchestration-script creation, while OpenAI's Codex / Agents SDK and Google's Gemini / Antigravity still expect developers to author orchestration logic explicitly. Anthropic favors fast exploration; OpenAI and Google suit stricter production governance.
  • Harness engineering shift: Developers move from writing task decomposition, spawn rules, state management, conflict resolution, and stop conditions to reviewing and tuning Claude-generated scripts—though high-security or compliance-heavy workflows still benefit from manual design and audit.
  • Reference

  • Docs: https://code.claude.com/docs/en/workflows

Tags

#claude-code#opus-4-8#dynamic-workflows#multi-agent#harness-engineering#anthropic#adversarial-verification#ai-coding

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980627