Software Factories and Harness Engineering

Our position. We hold that the harness matters more than the model: the model is a parameter, and the harness is the system the model runs in (The harness matters more than the model (opens in a new tab)). A harness is a list of things the model cannot yet do reliably on its own, so a good one is honest about exactly where it does not trust the model, no more and no less (What a harness is (opens in a new tab)).

Gates and validation have to scale faster than generation, because agents optimize to pass the gates, and weak gates turn high volume into unusable output (Dark Factory Patterns). The engineer’s job moves with that investment: you build the factory that builds the product.

A software factory is the standing system that results. It takes work in, routes it to agents, gates their output, releases what passes and learns from each run, in models that run from ticket-to-PR pipelines to dark factories and agent companies run as autonomous business units. Building a single agent belongs to methodologies and frameworks, and coordinating many agents belongs to multi-agent orchestration. The factory is the layer above both.

Start here

• Harness foundations — The components and control layers that turn a model into a dependable engineering agent.
• Harness maintenance — How to prevent configuration decay, remove stale machinery, and recognize harness anti-patterns.
• The enabling environment — The repository, tooling, feedback, and runtime conditions a software factory depends on.
• The software-factory pattern — How a delivery system takes intent to released software, and how to measure one.
• Factory roles and pipeline — How orchestrators, workers, and validators move work through a staged production pipeline.
• Agent observability - What to record about every agent run and every agent change so a factory can be debugged, audited and improved.

Related

• Dark Factory Patterns: the patterns of a factory that runs without a human approving each step.
• The software-factory pattern: how the standing system is measured and what running factories report.
• The enabling environment: the runtime and repository a harness depends on.
• Agent observability: the record that lets you debug and tighten a harness.
• Security and Runtime Isolation: the boundaries below the agent that a harness cannot replace.
• The thin-harness ecosystem: why a thin harness sheds scaffolding as each model release improves.

Resources

Agent Harnesses & Software Factories

• 8090 (opens in a new tab) - Full-SDLC software factory working upstream of code on requirements, architecture, and context, feeding agents a structured knowledge graph plus reverse-engineering agents for legacy systems.
• Agent Flywheel (opens in a new tab) - One-command installer that turns a fresh cloud server into an agentic coding environment with Claude Code, Codex, and Gemini preconfigured, plus a beads-based methodology guide.
• Amazon Bedrock AgentCore (opens in a new tab) - Framework- and model-agnostic AWS samples for Amazon Bedrock AgentCore, a suite of services for deploying and operating AI agents securely at scale.
• Archon (opens in a new tab) - Open-source harness builder that defines AI coding workflows in YAML across phases and runs them in parallel via git worktrees; ships 17 built-in workflows.
• Atomic (opens in a new tab) - MIT verifiable coding-agent runtime and CLI built on Pi’s CLI that runs structured workflows through explicit stages, review gates, and resumable checkpoints; model-agnostic.
• AutoForge (opens in a new tab) - Long-running autonomous coding agent on the Claude Agent SDK that builds full applications across sessions in a two-agent pattern, with a React progress UI.
• AutoMaker (opens in a new tab) - Open-source dev studio with a Kanban board that hands features to Claude-powered agents, which implement, test, and commit in isolated git worktrees.
• Awesome Harness Engineering (ai-boost) (opens in a new tab) - Curated list of harness-engineering resources: tools, patterns, evals, memory, MCP, permissions, observability, and orchestration.
• Awesome Harness Engineering (walkinglabs) (opens in a new tab) - Curated harness-engineering resources organized into foundations, context and memory, guardrails, agent files, evals, benchmarks, and reference runtimes.
• Background Agents (opens in a new tab) - Ona’s visual guide and white paper arguing background agents need cloud primitives, sandboxing, governance, triggers, and fleet coordination that a laptop cannot provide.
• Charlie Labs (opens in a new tab) - Hosted platform with a free tier and paid plans for always-on, Markdown-defined AI daemons that turn recurring engineering maintenance into verified, reviewable changes.
• Blitzy (opens in a new tab) - Enterprise autonomous coding service that runs 8-12 hour reasoning cycles across many agents to ship up to 3M lines per run, aimed at legacy modernization.
• Clipboard: Tickets to Pull Requests While You Sleep (opens in a new tab) - Clipboard Health engineer’s account of a pipeline that turns backlog tickets into reviewed pull requests overnight while the team sleeps.
• Evo (opens in a new tab) - Apache-2.0 autoresearch orchestrator that turns a codebase into an optimization loop, running tree search with parallel subagents in isolated git worktrees and sharing failure traces.
• Factory (opens in a new tab) - Agent-native development platform whose Droid agents run missions across the software lifecycle from a CLI, positioned as your software factory.
• Gas City (opens in a new tab) - MIT SDK in the Gas Town lineage for assembling, deploying, and supervising custom dark factories as composable agent packs, backed by the git-versioned Dolt store.
• Harness (opens in a new tab) - Claude Code meta-skill that turns a domain description into a specialized agent team and matching skills using six architecture patterns.
• Human (opens in a new tab) - AI software-factory pipeline that takes tickets, docs, designs, and analytics in and ships reviewed code out, with Claude doing the work and humans approving.
• Jig (opens in a new tab) - AI engineering workflow framework aimed at teams rather than solo developers.
• Mercury Agent (opens in a new tab) - AI agent with permission-hardened tools, token budgets, dual-channel access (CLI and Telegram), daemon mode with crash recovery, and multi-provider LLM fallback.
• Meta-Harness (Stanford IRIS Lab) (opens in a new tab) - Reference framework from the Stanford IRIS Lab paper for automated search over task-specific harnesses around a fixed base model, using Claude Code as proposer.
• Metaharness (opens in a new tab) - Python library and CLI for optimizing agentic coding harnesses through iterative improvement loops, with filesystem-backed run stores and provider-neutral backends.
• Omnigent (opens in a new tab) - Alpha-stage Apache-2.0 meta-harness that unifies multiple agent frameworks behind one orchestration layer, so harnesses swap or combine without rewriting, with multi-agent supervision.
• Ona (opens in a new tab) - Rebrand of Gitpod (September 2025) for running parallel software-engineering agents in ephemeral OS-isolated workspaces (managed cloud or your VPC), with SSO/OIDC, RBAC, and audit trails.
• Agent Factory (opens in a new tab) - GitHub’s public showcase of 100+ agentic workflows built on gh-aw, where markdown-defined agents compile into scoped GitHub Actions that triage issues and diagnose CI failures.
• SWE-CI (opens in a new tab) - Benchmark of 100 commit-pair tasks measuring long-term codebase maintenance, run as a CI loop between architect and programmer agents.
• Software Factory (Memo) (opens in a new tab) - Ona’s public Memo experiment building a product entirely through chained automations like Feature Planner, Builder, Verifiers, and Incident Responder, humans steering intent.
• Super Simple Software Factory (opens in a new tab) - MIT skill stamping agent-plus-code workflows into any repo, where Python owns sequencing and gates while agents run as bounded phases against typed envelopes.
• SwarmForge (opens in a new tab) - A tmux-based agent orchestration platform in Shell, coordinating coding-agent swarms with git worktree isolation; supports Claude, Copilot, Codex, and Grok backends.
• Symphony (opens in a new tab) - OpenAI’s experimental Elixir framework that turns project-board work items (e.g. Linear) into implementation runs by autonomous coding agents, verified through CI and PR reviews.
• claw-code (opens in a new tab) - Cleanroom Rust+Python reimplementation of a coding-agent harness with MCP integration, plugin system, hook pipelines, an interactive REPL, and OAuth API client.
• pi versus Claude Code (opens in a new tab) - Showcase of 6+ customized Pi Coding Agent (opens in a new tab) instances demonstrating agent teams, chains, and safety auditing; positioned as an open-source hedge against Claude Code.
• sgai (opens in a new tab) - Sandgarden’s goal-driven AI software factory in Go that runs multi-agent workflows locally in your repository, planning work as visual diagrams you approve before execution.

Memo Factory Automations

The complete automation set from Ona’s public Memo software factory, one YAML per standing agent.

• Memo Automation Auditor automation (opens in a new tab) - Weekly cron agent that audits the other automations’ performance and label hygiene, updates quality grades, and proposes prompt improvements.
• Memo Bug Fixer automation (opens in a new tab) - Two-hourly cron agent that picks the next open bug issue and fixes it autonomously after an environment health check.
• Memo Daily Metrics automation (opens in a new tab) - Cron automation from Ona’s Memo software factory that collects daily project stats into a JSON snapshot and opens a chore pull request.
• Memo Feature Builder automation (opens in a new tab) - Cron and merge-webhook agent that implements the next backlog feature autonomously and opens a pull request.
• Memo Feature Planner automation (opens in a new tab) - Hourly agent that triages unlabeled issues and decomposes the product spec into labeled, non-blocking backlog items.
• Memo Feedback Digest automation (opens in a new tab) - Hourly agent that reads new user feedback and usage events from Supabase and posts a categorized digest to Slack.
• Memo Incident Responder automation (opens in a new tab) - Hourly Sentry monitor that triages new production errors by severity into bug issues or immediate fixes.
• Memo Needs-Human Requeue automation (opens in a new tab) - Two-hourly agent that detects human replies on needs-human issues and removes the label so the planner re-triages.
• Memo PR Reviewer automation (opens in a new tab) - Webhook and half-hourly agent that reviews code, fixes CI failures, resolves review threads, and merges ready pull requests.
• Memo PR Shepherd automation (opens in a new tab) - Two-hourly agent that finds stalled, conflicted, or duplicate open pull requests and takes corrective action.
• Memo Performance Monitor automation (opens in a new tab) - Weekly production health check covering endpoint latency, Sentry error trends, and build size, filing issues on regressions.
• Memo Post-Merge Verifier automation (opens in a new tab) - Merge-webhook agent that waits out the production deploy and smoke-tests the live app after feature and fix PRs.
• Memo Product Improver automation (opens in a new tab) - Twice-weekly agent that reviews the live product and codebase for enhancement issues, gating high-risk ones for human approval.
• Memo Stale Issue Reviewer automation (opens in a new tab) - Four-hourly agent that detects issues stuck in progress without activity and resets them to backlog or escalates.
• Memo Tweet Drafter automation (opens in a new tab) - Scheduled automation from Ona’s Memo factory that drafts and posts build-in-public updates to X three times daily from repo metrics.
• Memo UI Verifier automation (opens in a new tab) - Merge-webhook agent that checks changed UI files against the design spec, statically and in the live app.
• Memo Weekly Recap automation (opens in a new tab) - Monday agent that summarizes the week’s merged PRs and metrics for the build-in-public audience.

Agent Companies & Business OS

• Augment Code (opens in a new tab) - Software agent company whose Cosmos coordinates humans, agents, code, tools, policy, and memory at the org level over a context engine with compounding memory.
• Companies.sh (opens in a new tab) - CLI installer for the Agent Companies open standard, importing portable AI company packages from GitHub, URLs, or local folders into orchestrators like Paperclip (paperclipai - source (opens in a new tab)).
• Multi-Agent Generator (opens in a new tab) - Low-code Python tool that turns natural-language descriptions into multi-agent teams across CrewAI, LangGraph, Agno, and ReAct, with five orchestration patterns and auto-generated test suites.
• Multica (opens in a new tab) - Open-source platform that assigns tasks to coding agents like human teammates, with squads under a leader for stable routing and autopilots scheduling recurring work.
• Paperclip (opens in a new tab) - Open-source orchestration for AI agent teams as autonomous business units with org charts, budgeting, governance gates, and audit logging across any agent type.
• Polsia (opens in a new tab) - Autonomous AI system that plans, codes, and markets a company through coordinated agents covering planning, development, marketing, and operations.
• The Agentic Startup (opens in a new tab) - Claude Code commands, skills, and agents framed as a virtual engineering team.