Research
Comparison Matrix
| System | Core abstraction | Workflow model | Roles/personas | State persistence | Spec format | Skill/capability format | Sandbox/isolation | Verification | Human checkpoints | Runtime portability | Copy into UH | Avoid in UH |
| System | Core abstraction | Workflow model | Roles/personas | State persistence | Spec format | Skill/capability format | Sandbox/isolation | Verification | Human checkpoints | Runtime portability | Copy into UH | Avoid in UH |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Specsafe | Issue/spec safety layer | Issue-driven enhancements | Not primary | GitHub issues | Issue/spec references | Integration-specific | Not primary | Spec safety implied | Issue review | IDE/CLI integration oriented | Traceability and integration roadmap | IDE-first lock-in |
| BMAD | Agile AI development method | Structured lifecycle workflows | Analyst, PM, Architect, Dev, QA, UX, etc. | Project/module files | Workflow artifacts | Agents/modules | Not core | Workflow reviews | Multiple approvals | Works through AI IDE installs | Role separation and lifecycle stages | Letting BMAD own the ontology |
| superpowers | Skills + coding methodology | Brainstorm → worktree → plan → TDD → review → finish | Skill-driven subagents/reviewers | Skill docs, plans, branches | Design/plan docs | Markdown skills | Git worktrees | Tests + reviews | Design approval, reviews, merge choice | Multiple agent hosts | Mandatory skill discovery and review loops | Single host/plugin assumptions |
| GSD | Context-engineered command loop | New project → discuss → plan → execute → verify | Fresh subagents | Project context and phase docs | Requirements/roadmaps/plans | Commands/prompts | Fresh contexts, not primarily FS | Verification loops | Phase gates | Multi-tool but Claude Code-centered | Context-rot mitigation and phase sizing | Permission model assumptions |
| matt-pocock/skills | Small engineering skills | Clarify → shared language → implement/debug/review | Skill-specific | CONTEXT.md and docs | Docs/tickets | Small reusable skills | Not primary | Engineering feedback loops | Clarification checkpoints | Agent-agnostic | Small practical skills and shared language | Over-process or verbose ceremonies |
| oh-my-openagent | Opinionated agent harness | Packaged workflows + hooks + routing | Multiple configured agents | Config, hooks, sessions | Harness config/docs | Agents, hooks, MCP tools | Runtime-dependent | Hooks/reviews | Workflow discipline | OpenCode-first, multi-model | Batteries-included defaults and hooks | Becoming a plugin distro |
| OpenSpec | Change artifact folder | Propose → apply → archive | Assistant commands | openspec/changes/... | Proposal, specs, design, tasks | Commands/skills | Not primary | Verify command and tasks | Propose/archive approvals | 25+ assistants | Artifact-guided lifecycle | Spec-only worldview |
| Pi / oh-my-pi | Terminal agent harness | Interactive/JSON/RPC/SDK; sessions and branches | Subagents | Session trees, config | Prompt/session artifacts | Skills/extensions/packages | Worktrees, FUSE overlay, ProjFS in oh-my-pi | Review command, LSP/tests | User steering and review | Multi-provider, embeddable | Adapter target and possible engine | Premature engine commitment |
| AgentFS | Agent filesystem | Init/mount/exec/run/sync | Not role-based | DB-backed filesystem | Filesystem state | MCP FS tools | Copy-on-write, overlay, sandbox | Inspect deltas/sync | Promotion by inspection | CLI/MCP | Sandbox backend design | Equating with git branches |
Matrix conclusions
- Ultimate Harness should be artifact-first, not runtime-first. OpenSpec, BMAD, superpowers, and GSD all succeed when important state survives outside chat.
- Runtime adapters must be structured, not shell-only. Pi, oh-my-pi, and oh-my-openagent show rich runtime state; adapters need lifecycle, logs, artifacts, diffs, and verification reporting.
- Skills should be portable procedures with metadata. superpowers and matt-pocock/skills show the value of small procedural units, but Ultimate Harness should allow richer machine-readable metadata over time.
- Sandboxing is a primary differentiator. AgentFS and git worktrees should both be modeled because they isolate different classes of risk.
- BMAD is a workflow profile family. It informs role mapping, but should not dominate the whole design.
Inspiration Systems
Ultimate Harness intentionally borrows from several systems while avoiding capture by any single one. The goal is not to reimplement BMAD, OpenSpec, GSD, or Pi; it is to extract portable patterns that survive across runtimes.
Adopt / Reject / Defer Decision Log
| Decision | Rationale | Source influence |