# Workflow Toolbox: A Systematic Construction Toolkit for Claude Code Multi-Agent Workflows > Workflow Toolbox is a multi-agent workflow construction framework for the Claude Code Workflow tool, providing seven validated orchestration patterns, writing scaffolds, and debugging tools to help developers build auditable and reproducible multi-agent workflow systems. - 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm) - 发布时间: 2026-06-13T13:47:29.000Z - 最近活动: 2026-06-13T13:58:52.367Z - 热度: 159.8 - 关键词: 多智能体, Claude Code, 工作流编排, TypeScript, 智能体协作, AI工作流, 模式库, 对抗性验证 - 页面链接: https://www.zingnex.cn/en/forum/thread/workflow-toolbox-claude-code - Canonical: https://www.zingnex.cn/forum/thread/workflow-toolbox-claude-code - Markdown 来源: floors_fallback --- ## Introduction / Main Floor: Workflow Toolbox: A Systematic Construction Toolkit for Claude Code Multi-Agent Workflows Workflow Toolbox is a multi-agent workflow construction framework for the Claude Code Workflow tool, providing seven validated orchestration patterns, writing scaffolds, and debugging tools to help developers build auditable and reproducible multi-agent workflow systems. ## Original Author and Source - Original Author/Maintainer: home-dev-lab - Source Platform: GitHub - Original Title: workflow-toolbox - Original Link: https://github.com/home-dev-lab/workflow-toolbox - Source Publish/Update Time: 2026-06-13T13:47:29Z --- ## Project Background and Core Positioning Claude Code's Workflow tool (research preview version) allows developers to orchestrate collaboration among multiple agents using JavaScript scripts, instead of relying on a single long conversation to complete all tasks. This architecture delegates control flows like loops, conditional judgments, and parallel execution to deterministic code—only the agent() calls at leaf nodes require real intelligent reasoning, and each agent runs in its own independent context. However, the Workflow tool itself only provides a basic platform. Developers need to implement a large number of repetitive mechanisms on their own: result validation after parallel execution, reasonable stop conditions, schema definition for result data, honest records of omissions or truncations, etc. Manually implementing these mechanisms easily leads to errors in the same subtle details, and a failed agent midway is almost indistinguishable from a successfully completed one. Workflow Toolbox is a systematic toolkit designed to address these pain points, analogizing the process of building multi-agent workflows to assembling Lego bricks. ## Molded Bricks (Pre-formed Blocks) `@workflow-toolbox` is a compile-time TypeScript library that provides seven tested and validated orchestration patterns, which can be combined using ordinary await/if/for statements. These patterns include: - **classifyAndAct**: Routing pattern—A classifier selects a category, then exactly one processor executes. - **fanOutAndSynthesize**: Parallel segmentation pattern—Independent agents run concurrently, and synthesis is triggered only after all results are received. - **adversarialVerification**: Adversarial validation pattern—Each claim faces refutation from independent verifiers; claims that fail the vote are retained and marked as unverified. - **generateAndFilter**: Generate-filter pattern—Extensive generation followed by a single filter based on clear criteria; rejected items are logged in real time. - **tournament**: Tournament pattern—Multiple attempts from different angles, scored by a panel of judges; the best ideas from the winner and runner-up are selected during synthesis. - **loopUntilDone**: Iterative optimization pattern with typed stop conditions—Supports four stop conditions: done/maxIterations/dryRounds/budgetFloor. - **planAndExecute**: Plan-execute pattern—The planner decomposes work into unpredictable subtasks; after the executor completes the subtasks, results are synthesized. ## Instruction Sheets (Guides) Claude Code skills (located in the plugin/ directory) teach Claude itself how to write, scaffold, and debug workflow scripts. These include: - **workflow-composer**: Writes runnable workflow scripts, handling file formats, pipeline vs parallel judgment, schema definition, deterministic rules, etc. - **toolkit-scaffold**: Generates a clean .workflow.ts skeleton connected to the selected @workflow-toolbox patterns. - **workflow-debugger**: Diagnoses completed or failed workflow runs from logs, analyzing agent failure reasons, pattern retry status, and whether safe recovery is possible. - **upgrade-canary**: Revalidates whether the Workflow runtime behavior matches the toolkit's dependencies after Claude Code or SDK upgrades. ## Finished Models (Ready-to-Use Workflows) Runnable workflows are provided as single-file .js artifacts in the toolkit/workflows/ directory. They can be directly executed by pointing the Workflow tool to their scriptPath without requiring a toolchain. This includes a complete family of development workflows (dev-plan → dev-implement → dev-review-fix, as well as dev-full which chains them together). ## Measured Results and Data-Driven Methodology The project's core philosophy is "Measured, not promised". Every claim is traced to logged production runs or public commits in this repository, including token counts, tool call counts, omission statistics for each agent, etc.—all of which can be audited post-hoc on the running machine. ## Quality Validation Results Adversarial review scans found 22 verified issues even after passing the code quality gates (tests + type checks + code linting). These include: the previous run fixed NUL bytes in files written by agents, and a rollback path that downgrades to a bare git reset --hard and silently retains incorrect merges when an agent self-reports an empty SHA. This proves that reviewers with fresh context can see issues that the author's context cannot.