CandleKeep
Anthropic

Dynamic Workflows in Claude Code: A Complete Guide to Multi-Agent Harnesses

by Research compilation — Anthropic blog + Claude Code documentation

workflows
Pages40
Formatmarkdown
ListedJune 3, 2026
UpdatedAugust 22, 2026
Subscribers50

About

Complete reference for Claude Code's dynamic workflows feature. Covers the JavaScript orchestration model, six workflow patterns (classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, loop-until-done), all major use cases (migrations, research, verification, triage, evals, model routing), prompting strategies, token budgets, and saving/sharing workflows. Based on Anthropic's official blog post by Thariq Shihipar and Sid Bidasaria.

Bring this into your next coding session

Install CandleKeep and your AI coding agent — Claude Code, Cursor, Codex, and more — can read Dynamic Workflows in Claude Code: A Complete Guide to Multi-Agent Harnesses directly. Cited, always current. Just ask, and the librarian pulls in the right book.

curl -fsSL https://getcandlekeep.com/install.sh | sh
40Chapters
88Topics
40Pages

Preview

Dynamic Workflows in Claude Code: A Complete Guide to Multi-Agent Harnesses

Load this book when:

  • You are deciding whether a task should run as a Claude Code dynamic workflow instead of a single turn.
  • You are choosing which of the six core workflow patterns (classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, loop-until-done) fits a task.
  • You are diagnosing agentic laziness, self-preferential bias, or goal drift in a long or parallel task.
  • You are wiring ultracode, /goal, or /loop into a workflow and need the exact trigger and stop-condition mechanics.
  • You are routing subtasks across Haiku/Sonnet/Opus or setting a token budget for an orchestration.
  • You want to understand Loop Engineering — designing systems that prompt agents, not prompting agents yourself.
  • You need to build a durable agent loop with automations, worktrees, skills, sub-agents, and memory.
  • You are choosing between a dynamic workflow (Claude writes the JS) and a deterministic workflow (you write it once and commit it).
  • You need a multi-hour job (bulk API pull, large scrape, long batch export) to survive session restarts, background-task kills, and an absent user.

Author: Research compilation — Anthropic blog + Claude Code documentation + community research (Addy Osmani, Forward Future Loop Library, Boris Cherny) Source: "A harness for every task" — Anthropic Blog, 2026 (Thariq Shihipar & Sid Bidasaria); "Loop Engineering" — Addy Osmani, June 2026; Forward Future Loop Library, June 2026 Target: Claude Code users who want to orchestrate multi-agent workflows for complex tasks Last updated: 2026-08-22 Audience & governance: Marketplace (public) reference book — optimize for circulation and attribution (per Writing Books for AI Agents, Ch 13). Valid as of: 2026-06-19, Claude Code v2.1.154+ (dynamic workflows GA). Refresh cadence: review every release that bumps subagent caps, model lineup, or the /goal//loop/ultracode surface (tactical/volatile domain — short cadence).


Add to library to read more

Table of Contents

Dynamic Workflows in Claude Code: A Complete Guide to Multi-Agent Harnesses
Chapter 0 — Scope & Currency
Chapter 1 — The Problem Workflows Solve
Chapter 2 — Decision Matrix
Chapter 3 — Dynamic vs Static Workflows

Chapter 11 — Migrations and Refactors
Chapter 12 — Deep Research
Chapter 13 — Deep Verification
Chapter 14 — Sorting at Scale
Chapter 15 — Memory and Rule Adherence
Chapter 16 — Root-Cause Investigation
Chapter 17 — Triaging at Scale
Chapter 18 — Exploration and Taste
Chapter 19 — Evals

Chapter 21 — Triggering Workflows
Chapter 22 — Prompting Strategies
Chapter 23 — Token Budgets
Chapter 24 — Saving and Sharing Workflows

Chapter 26 — Pattern Selection Guide
Chapter 27 — Use Case Quick Reference
Chapter 28 — Failure Mode Reference
Chapter 29 — Example Prompt Gallery

Chapter 34 — Loop Maturity Ladder

Add to Library

Free · Live updates included

50 readers subscribed