Skip to content

00 - Vision

Agent Orchestrator is a Harness Engineering control plane for agent-first software delivery.

OpenAI recently used the term Harness Engineering to describe an engineering style where humans focus less on hand-writing code and more on designing environments, specifying intent, and building feedback loops that let agents do reliable work. This project adopts that framing directly.

What We Are Building

  • A local-first control plane that abstracts Claude Code, Codex, OpenCode, Gemini CLI, and other shell-based tools behind one resource model
  • A declarative harness where workflows, agents, triggers, secrets, policies, and observability live as versioned repository assets
  • A long-running execution runtime that can keep plan -> implement -> test -> review -> fix loops alive for hours or days

Why a Control Plane?

A serious engineering harness does more than invoke an agent once. It decides:

  • which agent should run
  • in which workspace
  • on which workflow step
  • with which guardrails and secrets
  • under which retry, pause, or recovery policy
  • and how results are turned into auditable state, logs, and downstream actions

That is the role of this project.

Core Beliefs

  • Humans steer, agents execute: people set goals, acceptance criteria, and constraints; the system coordinates execution.
  • The repository is the system of record: manifests, docs, skills, QA artifacts, and policies should be versioned and discoverable in-repo.
  • Guardrails matter more than heroic prompting: workflows, invariants, triggers, and recovery paths compound more than one-off prompts.
  • Any shell-based agent should be portable: the control plane should outlive any single model or vendor.
  • Long-running software delivery loops must be observable: durable state, logs, events, and traces are first-class requirements.

What This Is Not

  • Not a chat wrapper around one model
  • Not just a workflow DSL
  • Not a generic CI/CD replacement for every enterprise use case

The ambition is narrower and more opinionated: become the control plane for agent-native engineering systems.

Product Direction

  • Turn ad hoc agent usage into reusable workflow assets
  • Let teams institutionalize QA, fix, review, governance, migration, and self-bootstrap loops
  • Provide a stable harness layer above rapidly changing models and coding shells

Continue Reading