Dimensions
25 engineering dimensions
The RUBRIC-v3 schema evaluates each framework across 25 engineering dimensions. Each dimension page lists which frameworks have it (and which don't), sorted by stars. The numbers reveal which engineering rigor is common, which is rare, and which is unique.
Whether the framework enforces test-driven development — and how structurally: from prompt suggestion to compiled invariant.
Frameworks spawning N concurrent agents (swarms, hierarchies, director/coder splits).
Per-role model selection — haiku for fast, opus for careful, cross-vendor for cost arbitrage.
Persistent, queryable record of agent actions for compliance and replay.
Ability to re-run a past session deterministically from its audit log.
Automatic recovery from agent loops, context exhaustion, or tool failures.
Each task runs in an isolated git worktree, enabling true parallel development.
Explicit strategy for when context window fills — PreCompact hooks, summarization, handoff.
State transfer between agent sessions, enabling long-horizon tasks.
Dedicated CLI binary (vs. just being a config file bundle).
Bundles one or more MCP (Model Context Protocol) server tools.
Ships a local web/desktop dashboard (not just a terminal interface).
How the framework isolates tasks — git worktrees, Docker containers, sandboxes, PTYs.
Structured multi-turn interactions where outputs feed into subsequent prompts.
How the framework encodes specifications — Markdown, YAML, XML, symbolic notation, DSL.
Framework commits to git automatically after task completion.
Framework opens a pull request automatically.
Framework merges the PR automatically when CI passes.
Outputs stream in real-time rather than batching at completion.
Persistence strategy: file-based, SQLite, vector, HTTP-proxy, episodic.
How broadly the framework targets different AI agents (single-tool vs. multi-runtime).
How the framework is packaged — Claude plugin, npm, pip, git clone, YAML bundle.
Human-in-the-loop checkpoints before the agent proceeds to the next phase.
Named phases of work (plan, spec, implement, review, verify) vs. a single-phase agent.
How multiple agents relate — queen+workers, director/coder, peer-to-peer, pipeline.