open-dynamic-workflows – skill Agent & Tự động hóa cho agent AI

Open Dynamic Workflows

Overview

Open Dynamic Workflows (ODW) is an open-source dynamic multi-agent workflow engine for AI coding agents such as OpenCode, Codex, Antigravity, and VS Code. It lets you plan a task, orchestrate multiple agents working in parallel, and adversarially verify their output before it lands. ODW ships a Codex/Antigravity skill folder (SKILL.md plus a daemon bridge) and an OpenCode plugin, and it is bring-your-own-model (Anthropic, OpenAI-compatible, or Ollama). This skill is adapted from the community project at Suraj1235/open-dynamic-workflows.

When to Use This Skill

• Use when you need to decompose a coding task into independent subtasks and run multiple agents in parallel.
• Use when working across more than one AI coding tool (OpenCode, Codex, Antigravity, VS Code) and want a single orchestration layer.
• Use when the user asks for adversarial review or verification of agent-generated changes before merging.

How It Works

Step 1: Plan

ODW takes a high-level goal and produces a dynamic workflow graph of subtasks, identifying which can run in parallel and which have dependencies.

Step 2: Orchestrate

The engine dispatches subtasks to parallel agents through the OpenCode plugin or the Codex/Antigravity daemon bridge, using your configured model provider (Anthropic, OpenAI-compatible, or Ollama).

Step 3: Adversarially Verify

Completed work is routed through an adversarial verification pass that challenges the output before results are synthesized and returned.

Examples

Example 1: Run a parallel workflow

ODW requires a separately installed upstream runtime; this single-file skill does not bundle its daemon, bridge, plugin or workflow scripts. Inspect the full upstream tree and pin a reviewed immutable revision before installation. Dependency installation can execute lifecycle scripts; inspect those scripts and the lockfile first, and use an isolated environment. The CLI is

odw-daemon — run it as npm run odw -- <args> from inside the repo, or as

npx odw-daemon <args> / a global odw-daemon if you link the bin.

# Configure your model provider (bring-your-own-model)
export ANTHROPIC_API_KEY=...        # or an OpenAI-compatible / Ollama endpoint

# One-time setup: generate ~/.odw/config.json
npm run setup

# Start the local workflow daemon (once)
npm run odw -- start

# Plan, orchestrate, and verify a task across parallel agents
npm run odw -- run --prompt "refactor the auth module and add tests"

Example 2: Use the Codex/Antigravity skill bridge

# ODW ships a SKILL.md + daemon bridge consumed by Codex / Antigravity.
# Start the daemon, then run a saved orchestration script through it:
npm run odw -- start
npm run odw -- run --script examples/workflows/studio-prime.workflow.js --cwd .

Best Practices

• ✅ Scope each subtask so agents can run without shared state.
• ✅ Keep the adversarial verification pass enabled before merging agent output.
• ❌ Don't run interdependent subtasks in parallel without declaring their dependencies.
• ❌ Don't commit provider API keys; use environment variables or a secrets manager.

Limitations

• This skill does not replace environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, or safety boundaries are missing.

Security & Safety Notes

• ODW executes agent-generated code and shell commands. A local directory alone is not isolation: use an authorized sandbox with minimal filesystem, credential and network access. Bound concurrency, model spending and permitted commands; adversarial review does not grant merge or production authority.
• Model provider credentials (Anthropic / OpenAI-compatible / Ollama) must be supplied via environment variables, never committed to source.
• Review adversarial-verification output before applying changes to a production branch.

Common Pitfalls

• Problem: Parallel agents collide on the same files.

Solution: Give each subtask exclusive file/module ownership and run conflicting tasks sequentially.

Related Skills

• @multi-agent-orchestration - When coordinating multiple agents on one goal.
• @code-review - How adversarial verification complements human review.