Skip to content
/
Phase B Batch 15

Batch 15 — Explicit "Harness" Frameworks

Batch 15 — Explicit "Harness" Frameworks

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
chachamaru-cc-harness 1,611 claude-plugin yes (Go harness) none hierarchical (worktree workers) yes (haiku for hook review) A
revfactory-cc-harness 98 standalone-repo no none parallel-fan-out (A/B experiment) no A
chapzin-codex-harness-mcp 7 mcp-server no none none (MCP infrastructure) no A
adrielp-ai-engineering-harness 16 cli-tool (Deno ai-harness) yes (install-time only) none task-decomposition-tree no A
bquast-simple-agent-harness 0 bash-script-bundle no none none (single agent, cron) no A
ldzhouquan-agent-harness-skill 9 skill-pack no none sequential (Reflexion Loop) no A
revfactory-harness 3,552 claude-plugin no none hierarchical (generated, all-opus) yes (all-opus mandate) A
openharness 13,115 standalone-repo (pip) yes (Python oh) terminal-tui + web-dashboard swarm (subprocess) yes (provider-agnostic) A
hermes-agent 167,908 standalone-repo (pip) yes (Python hermes) terminal-tui + web-dashboard parallel-fan-out yes (provider-agnostic) A
ecc-affaan 193,231 claude-plugin yes (Rust ecc2 alpha + npm) desktop-app (Tkinter) parallel-fan-out + worktree yes (haiku/sonnet/opus routing) A

Archetype Placement

slug Primary archetype Closest seed
chachamaru-cc-harness Archetype 1 (skills behavioral) + Go binary superpowers
revfactory-cc-harness Unique: measurement apparatus agent-os
chapzin-codex-harness-mcp Archetype 3 (MCP-anchored) ccmemory
adrielp-ai-engineering-harness Archetype 2 (mirror cmd+skill, multi-tool) openspec
bquast-simple-agent-harness Archetype 4 (markdown scaffold) extreme minimal agent-os
ldzhouquan-agent-harness-skill Archetype 1 (skills behavioral, Iron Laws) superpowers
revfactory-harness Unique: meta-factory (generates harnesses) BMAD-METHOD
openharness Archetype 5-adjacent (standalone Python runtime) claude-flow / kiro
hermes-agent Unique: self-improving personal agent claude-flow
ecc-affaan Archetype 1 dominant, expanding toward Archetype 5 superpowers + claude-flow

Intra-batch patterns

All 10 frameworks explicitly call themselves "harnesses" but mean radically different things: Chachamaru and ECC mean "delivery loop harness" (wraps agent work in a plan/verify cycle), revfactory-harness means "generator of harnesses" (meta-level), chapzin means "control-plane harness" (state and governance for an existing agent), bquast means "minimal autonomous agent harness" (the agent runtime itself), OpenHarness and Hermes mean "agent execution harness" (the infrastructure layer the agent runs in), and ldzhouquan means "engineering methodology harness" (behavioral rules). The word "harness" does real semantic work: it always connotes a constraint layer that keeps work within bounds — but the boundary sits at different levels: delivery loop, execution runtime, behavioral methodology, state tracking, or team architecture generation.

The star distribution is extreme: ECC (193k) and Hermes (168k) dominate; the remaining 8 sum to under 20k. The two standalones (OpenHarness, Hermes) and the ECC super-framework are production-grade deployed applications; the rest range from research (revfactory-cc-harness) to minimal (bquast).

Korean/Japanese multilingual documentation appears in 4 of 10 frameworks (chachamaru, revfactory-cc-harness, revfactory-harness, ldzhouquan), suggesting strong East Asian engagement with Claude Code harness engineering.

Most interesting finds

  1. revfactory-harness (3,552 stars): The only framework in the entire catalog that is a harness generator rather than a harness. Its explicit L3/L2/L1 ecosystem taxonomy — naming ECC as L2 and Archon as a co-existing L3 — is the most sophisticated meta-commentary on framework design found in any batch. The all-opus mandate for all generated agents is a bold and expensive opinion worth testing.

  2. revfactory-cc-harness (98 stars): The only framework that provides quantitative proof of harness value (79.3 vs 49.5 quality score, 15-task A/B experiment). No other framework in the entire research corpus provides controlled empirical evidence. The finding that harness effectiveness scales with task complexity (+36.2 points at Expert level) is a key design validation that justifies the entire research direction.

Items written as Tier C

None. All 10 frameworks had sufficient public material for complete 11-file reports.

Cross-references discovered

  1. ECC ↔ revfactory-harness: revfactory-harness README explicitly names ECC as an "L2 Cross-Harness Workflow" co-existing tool, positioning Harness as the L3 Meta-Factory above it. ECC's README in turn names revfactory-harness as a companion in its ecosystem documentation.

  2. chachamaru-cc-harness ↔ ECC: Chachamaru's README explicitly warns not to inherit support claims from "Superpowers, Hermes Agent, or any other project" — naming ECC's predecessor community. The breezing/parallel execution design resembles ECC's cascade method.

  3. hermes-agent → ECC: ECC's autonomous-agent-harness skill explicitly lists "Hermes" as one of the standalone agent frameworks it replaces ("Spawn isolated subagents... Replaces standalone agent frameworks (Hermes, AutoGPT)").

  4. openharness ↔ hermes-agent: Both are standalone Python agent runtimes with TUI + messaging gateway + multi-provider support, released around the same time (OpenHarness v0.1.x, Hermes v0.14.x). OpenHarness README mentions OpenClaw and nanobot (OpenClaw is Hermes' predecessor). Independent parallel development or conscious positioning.

  5. revfactory-cc-harness (A/B experiment): The experiment project validates the same .claude/ pre-configuration pattern that chachamaru-cc-harness, adrielp-ai-engineering-harness, and ldzhouquan-agent-harness-skill all implement — providing empirical backing for the entire harness-as-context-engineering school.

  6. adrielp-ai-engineering-harness ↔ hermes-agent: The TDD skill in adrielp's harness and the TDD skill in hermes-agent's software-development category are nearly identical in wording ("Anti-Pattern: Horizontal Slices," "WRONG/RIGHT visual") — likely drawn from the same origin (possibly the Anthropic engineering blog or a shared community reference).