Title here
Summary here
What is Unbound Force?
The AI Agent Swarm for Software Engineering
Unbound Force is a set of purpose-built agent personas designed to work as a coordinated software engineering team. Themed as a superhero squad, each persona fills a distinct role — Product Owner, Developer, Tester, PR Reviewer, and Manager — and operates inside your development environment as an AI-powered collaborator.
But these are not just instruction files. Each hero can include LSP servers, MCP servers, tooling, tasks, commands, plugins, and other technologies that enable them to do their job. They are designed to be the pinnacle archetype of their role.
The Stack
Unbound Force is built on four complementary tools that form a layered stack:
| Layer | Tool | What It Does |
|---|---|---|
| CLI | uf | Project scaffolding, environment setup, health checks, configuration, sandboxed execution, LLM gateway. |
| Agent | OpenCode | The AI coding environment where you interact, write code, and run commands. The personas run inside OpenCode. |
| Planning | Speckit (spec-kit) | A specification pipeline that turns ideas into structured specs, plans, and tasks before implementation begins. |
| Coordination | Replicator | Multi-agent coordination: parallel workers, git-backed tracking, file reservations, and semantic memory. Single Go binary. |
Each tool is independently useful, but they compose into the full Unbound Force workflow: scaffold with uf, plan with Speckit, execute with OpenCode, coordinate with Replicator.
The Heroes
The swarm consists of five personas, each representing the pinnacle archetype of their role:
- Muti-Mind — Product Owner. The Vision Keeper and Prioritization Engine.
- Cobalt-Crush — Developer. The Engineering Core and Adaptive Implementation Engine.
- Gaze — Tester. The Quality Sentinel and Predictive Validation Engine.
- The Divisor — PR Reviewer Council. The Architectural Conscience and Code Integrity Guardian.
- Mx F — Manager. The Flow Facilitator and Continuous Improvement Coach.
Together, they cover the full software development lifecycle – from requirements and planning through implementation, testing, review, acceptance, and reflection. With Dewey configured, the swarm runs autonomously from define through review – the human provides a short seed (1-2 sentences of intent) and reviews the completed increment. Without Dewey, the human drives the define stage and the swarm runs implementation through review.
Get Started
Ready to dive in? Start with the Quick Start guide to install the tools, then pick the guide for your role:
- Developer / Engineer – Daily workflow, Speckit pipeline, Replicator coordination, convention packs
- Tester – Gaze quality analysis, CRAP scores, coverage ratchets, CI integration
- Product Owner – Muti-Mind backlog management, priority scoring, acceptance decisions
- Product Manager – Mx F metrics, dashboards, coaching, retrospectives
- Common Workflows – The
/unleashautonomous pipeline,/finaleshipping workflow, manual feature flows, bug fixes, and code reviews - Hero Artifacts – Inter-hero communication: envelope format, artifact types, and lifecycle data flow
- Knowledge Retrieval with Dewey – Install and configure Dewey for semantic search across your repositories
- Constitution – The 4 core principles that govern all heroes and the governance model
Design Philosophy
An AI agent is a model plus a harness — the surrounding system of context, controls, and feedback loops that shapes what the model produces. The model provides capability; the harness provides direction. Unbound Force invests heavily in the harness because the same model with a better harness consistently outperforms a better model with a weaker harness.
The system delivers context to agents through three tiers. Static documentation (AGENTS.md) provides project structure and conventions at session start. Versioned rules ( convention packs) encode coding standards as numbered, severity-classified rules that are portable across projects. Dynamic semantic memory ( Dewey) provides searchable context from prior sessions, GitHub issues, and web documentation — adapting over time as the knowledge base grows.
Quality is enforced through layered feedback: computational checks first (tests, linters, Gaze static analysis), then inferential review ( the Divisor Council — multiple specialized agents evaluating from distinct quality dimensions). The doer and the judge are structurally separated — review agents cannot modify files, so findings must go through the implementation path with full visibility. This layered approach means fast, cheap, deterministic checks catch the obvious issues before slower, semantic review agents spend cycles on deeper analysis.