Orchestrate parallel AI agents with adversarial verification

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 plans a high-level goal into a dynamic workflow graph of subtasks, identifying which can run in parallel and which have dependencies; orchestrates by dispatching those subtasks to parallel agents through an OpenCode plugin or a Codex/Antigravity daemon bridge, using a bring-your-own-model provider (Anthropic, an OpenAI-compatible endpoint, or Ollama); and then routes completed work through an adversarial verification pass that challenges the output before results are synthesized and returned. This skill is adapted from the community project Suraj1235/open-dynamic-workflows.

Use it to decompose a coding task into independent subtasks and run multiple agents on them in parallel, when working across more than one AI coding tool (OpenCode, Codex, Antigravity, VS Code) and wanting a single orchestration layer, or when the user asks for adversarial review or verification of agent-generated changes before merging. Scope each subtask so agents can run without shared state, and declare dependencies rather than running interdependent subtasks in parallel; give each subtask exclusive file or module ownership to avoid parallel agents colliding on the same files, running conflicting tasks sequentially instead. Keep the adversarial verification pass enabled before merging, and review its output before applying changes to a production branch - ODW executes agent-generated code and shell commands, so it should run only in an authorized, local, or sandboxed environment. It does not replace environment-specific validation, testing, or expert review.

ODW is installed from source (clone the repo, then npm install); its CLI is odw-daemon, run as npm run odw -- <args> from inside the repo, or as npx odw-daemon <args> or a global odw-daemon if the bin is linked. Input is a natural-language prompt describing the goal, or a saved orchestration script; output is the synthesized, adversarially-verified result of the parallel agent runs. A typical run:

### 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"

Model provider credentials must be supplied via environment variables, never committed to source.

Ships a Codex/Antigravity skill folder (SKILL.md plus a daemon bridge) and an OpenCode plugin, and works with VS Code as one of its supported coding-agent surfaces. It is bring-your-own-model: Anthropic, any OpenAI-compatible endpoint, or Ollama. It pairs with a general multi-agent-orchestration skill for coordinating multiple agents toward one goal, and complements human code review by handling the adversarial-verification pass before a human reviewer sees the change.

Developers coordinating multiple AI coding agents - across OpenCode, Codex, Antigravity, or VS Code - who want to decompose a task into parallel subtasks, run them under a chosen model provider, and adversarially verify the combined output before it merges.