AgentPipeline

Let Codex/Claude Code run long task chains unattended with real branching

Problem

People running deep-learning or simulation projects have to read Codex's last output and manually decide the next step, staying glued to the terminal so their nightly quota goes unused. When someone asked for an unattended continuous-execution mode, the only answers were setting alarms or cron queues, which only work when every subtask is known upfront.

Solution

A local task orchestrator: describe the chain as a small DAG or rule set ('if the previous output matches this, run A, else B'), automatically feed each step's stdout and artifacts into the next agent session, and push a notification to your phone when a run fails or needs human judgment.

Users

Indie developers and researchers running long AI-agent experiments, simulations, and data pipelines

MVP

YAML-defined task chains with dependencies; wrapper around CLI agents capturing output; conditional branching and retries; push notifications on failure or pending decision; per-run logs and artifact archive

Revenue

Open-source core plus a Pro subscription for multi-machine parallelism, notification channels, and cloud run history, or a per-seat one-time license

Why now

Agent CLIs can now run autonomously for hours, but there's no lightweight orchestration layer connecting multi-step decisions, and only developers living this workflow can design it well