Agent Harness System

An open-source multi-agent harness that runs unattended — orchestrator-worker topology with per-agent context budgeting and cycle detection.

Most multi-agent demos loop forever or bankrupt you on tokens the moment you scale past a handful of agents. This harness is built to run, not just to demonstrate. It coordinates many concurrent LLM sub-agents across multiple frontier models while keeping cost and reliability under control.

What it does

• Hierarchical orchestrator-worker topology — a planner agent decomposes a goal into a task graph; a router assigns sub-tasks to workers by model capability.
• Per-agent context budgeting — each worker gets a scoped scratchpad view and a token budget, preventing the cost explosion that naive shared-context multi-agent systems hit.
• Dependency-cycle detection on the task graph — stops agent loops before they start. The difference between a demo and a system that runs overnight.
• Max-iteration caps — every agent has a hard stop.
• 16 agents, 32 commands — a working command surface, not a stub.

Why this exists

I run a personal multi-agent harness (Swarm OS) coordinating 50+ concurrent agents across 5 frontier models. The reliability primitives here — context budgeting, cycle detection, iteration caps — are the ones that turned it from "cool for 10 minutes" into "runs unattended and cuts multi-step task latency ~40% versus a single-agent baseline."

This repo is the open-sourceable core of that work.

Repo layout

README.md            this file
UNIVERSAL_PROMPT.md  base prompt contract for all agents
rules/               agent rules / behavior specs
scripts/             orchestration + runner scripts
CHANGELOG.md
CONTRIBUTING.md
DECISIONS.md         architectural decision records
INTEGRATION_TEST.md
LICENSE

Design principles

1. Scope before you scale — a 50-agent system with shared context is a 50× cost multiplier. Scoped scratchpads break that.
2. Fail bounded — every agent has a max-iteration cap and a cycle detector. No infinite loops reach the API.
3. Capability-aware routing — cheap models handle cheap sub-tasks; frontier models handle the hard graph nodes.

Status

Working harness, used as the backbone of personal multi-agent workloads. Open-sourced for reuse.

License

MIT