Now open source

agentic-harness

Open-source pipeline orchestration for AI-driven software delivery.

MIT Licensed Agent-Agnostic Production-Tested
View on GitHub Get Started
01Issue
02Spec
03Design
04Build
05QA
06Done
How It Works

Labels drive the
state machine.

GitHub issue labels define pipeline state. When an issue is labeled pipeline/spec, the harness spawns a fresh agent to execute that stage.

Each stage transition advances the label. A fresh agent picks up the next stage with clean context. No shared memory, no accumulated drift — just the issue, the repo, and the stage contract.

If an agent crashes, restart it. The label is still there. The pipeline resumes.

Labels as State

Pipeline state lives on the issue, not in memory. Labels are the single source of truth for where work stands.

Fresh Agent Per Stage

Every stage spawns a new agent with clean context. No prompt drift, no accumulated hallucinations.

Stateless by Design

Agents don't share memory. Stage contracts define inputs and outputs. The repo is the shared state.

Crash-Resilient

If a process dies, the label persists. Restart the harness and it picks up exactly where it left off.

Design Principles

Built for reliability,
not complexity

Five architectural decisions that make the harness simple to run, easy to extend, and hard to break.

Labels as State Machine

Pipeline state is encoded in GitHub labels. No external database, no orchestration server. Labels are durable, inspectable, and version-controlled.

Stateless Agents

Each agent starts clean. No shared memory between stages. Stage contracts define exactly what goes in and what comes out.

Agent-Agnostic

Swap in Claude, GPT, Codex, Gemini, or your own model. The harness orchestrates the pipeline — it doesn't care what agent runs each stage.

Multi-Pipeline

Run multiple issues through the pipeline concurrently. Each issue tracks its own state independently. Parallelism is built in.

Crash-Resilient

Process dies? Label persists. Restart the harness and it resumes from exactly where it stopped. No data loss, no manual recovery.

Production Results

Tested. Deployed.
QA passed.

Real applications built end-to-end by the harness — from issue creation to live deployment.

Counter App

Single file, vanilla JS

~20m Time
< $1 Cost
Deployed, QA passed

Journaling App

React SPA, 13 components

~25m Time
~$5.50 Cost
Deployed, QA passed

Iterative Enhancement

Add markdown to existing app

~15m Time
~$2 Cost
Extended codebase, QA passed
What's Next

From open source
to enterprise.

agentic-harness is the pipeline engine. Foundary is the governed control plane built on top of it — adding the approval gates, audit trails, and policy controls that enterprises require.

Learn about Foundary Book a Demo

  • Durable Approval Gates

    Every stage transition requires explicit sign-off. Recorded, timestamped, auditable.

  • Artifact Capture

    Code diffs, test results, deploy logs — automatically attached to every pipeline run.

  • Runtime Policy

    Control which tools, runtimes, and MCP servers agents can access — by stage and team.

  • Hybrid Deployment

    Hosted control plane, customer-run workers. Code and secrets never leave your boundary.

  • Operator Dashboards

    See run history, blocked work, approvals, and outcomes across repos and teams.

Get Started

Clone. Configure.
Run.

Three commands to go from zero to a running pipeline. The harness watches your GitHub repo for labeled issues and drives them through the full SDLC.

GitHub Documentation 
# Clone the repository
git clone https://github.com/ascendantventures/agentic-harness-oss.git
cd agentic-harness-oss

# Install dependencies
npm install

# Configure your environment
cp .env.example .env
# Edit .env with your GitHub token and agent config

# Start the harness
npm start