Agent Pipeline.

01 · The problem

Agentic work eats context.

Common agentic tasks consume time and context that belong to the main agent loop, and exploring divergent design concepts one at a time is slow. Agent Pipeline offloads both: orchestrated agents run the routine work and the explorations in parallel, away from the loop that matters.

02 · The approach

Everything is a file.

Agents are just .md files. Pipelines are schema-validated YAML with DAG-planned parallelism, and each run executes in an isolated git worktree with atomic commits per stage. A handover file passes context from one agent to the next while humans stay in the loop through a live terminal UI and GitHub pull requests.

$ npm install -g agent-pipeline
$ agent-pipeline init
✓ scaffolded 3 example pipelines + agent definitions
$ agent-pipeline run front-end-parallel-example
◆ plan     DAG resolved — parallel stages queued
✓ run      isolated worktree · atomic commits per stage
█

03 · Capabilities

Six systems, one CLI.

• 01

  Pipeline Orchestration

  DAG-planned parallelism with conditional gating, per-stage retries, and automated execution flow.

• 02

  Git Workflow Automation

  Branch and worktree isolation, atomic commits per stage, and automated GitHub PR creation via CLI integration.

• 03

  Multi-Runtime Flexibility

  Run stages on Claude Code, Codex, Gemini, or Pi Agent, and mix models per stage for cost optimization.

• 04

  Live Terminal UI

  Ink-powered real-time visualization with status badges, streaming output, and interactive history browser.

• 05

  State & Context Management

  Persistent run history with filesystem-based handover enabling seamless communication between stages.

• 06

  YAML-First Configuration

  Schema-validated pipelines with filesystem-based stage handover and customizable commit messages.

05 · Under the hood

Built on plain things.

Runtimes

Claude Code · Codex · Gemini · Pi Agent

Requires

Node 18+ · Git

Distribution

npm · open source

Pull requests

GitHub CLI