The developers shipping 10x faster aren't writing better prompts - they're running better infrastructure around their agents. These four repos cover the whole stack: turning specs into code, spinning up agent teams, bolting memory and security onto your agent, and making destructive commands structurally impossible. Combined star count north of 330K. All open source.
Links & Resources
- Spec Kit (117K stars)
- Spec Kit docs
- Harness (3.6K stars)
- ECC - Everything Claude Code (210K+ stars)
- AgentShield
- Agent Governance Toolkit (3.6K stars)
1. Spec Kit - GitHub''s Spec-Driven Development Toolkit
117K stars. MIT. From GitHub itself.
Instead of prompting your agent from scratch every time, you write an executable spec and the agent turns it into a plan, task list, and working code. The spec becomes the source of truth.
It works through slash commands that run in sequence:
| Command | What it does |
|---|---|
/speckit.constitution | Set project principles |
/speckit.specify | Define what you''re building |
/speckit.plan | Create the technical plan |
/speckit.tasks | Generate an actionable task list |
/speckit.implement | Execute all tasks |
/speckit.analyze | Cross-artifact consistency check |
Works with 30+ AI coding agents - Claude Code, Copilot, Gemini CLI, Codex, Cursor. Install:
uv tool install specify-cli --from git+https://github.com/github/spec-kit.git
specify init my-project --integration claude
2. Harness - A Factory That Designs Whole Agent Teams
3.6K stars. Apache-2.0.
Say "build a harness for this project" in Claude Code and Harness reads your domain, then generates a full team of specialized agents plus their skills - written into .claude/agents/ and .claude/skills/. A meta-skill that builds agent teams rather than being one.
It picks from six team patterns:
| Pattern | Use case |
|---|---|
| Pipeline | Sequential dependent tasks |
| Fan-out/Fan-in | Parallel independent tasks |
| Expert Pool | Context-dependent selective invocation |
| Producer-Reviewer | Generation followed by review |
| Supervisor | Central agent distributing tasks |
| Hierarchical Delegation | Top-down recursive delegation |
The author''s A/B test (n=15) reported +60% average quality with a 15/15 win rate - flagged as author-measured, third-party replications pending.
/plugin marketplace add revfactory/harness
/plugin install harness@harness
Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.
3. ECC (Everything Claude Code) - The Operator System
210K+ stars. MIT.
One of the most-starred repos in the Claude Code ecosystem. A complete operator layer that bolts skills, persistent memory, continuous learning, and security scanning onto your agent. Runs across Claude Code, Codex, Cursor, OpenCode, Gemini, Zed, and GitHub Copilot.
Ships with 67 agents, 277 skills, and 93 command shims.
| Layer | What it does |
|---|---|
| Skills | The primary workflow surface |
| Memory persistence | Hooks that save/load context across sessions |
| Instincts | Confidence-scored learning you can evolve into skills |
| AgentShield | Security auditor - 1282 tests, 102 static-analysis rules |
| Research-first | A search-first skill that researches before it codes |
AgentShield scans your CLAUDE.md, settings.json, MCP configs, hooks, and agent definitions:
npx ecc-agentshield scan # quick scan, no install
npx ecc-agentshield scan --fix # auto-fix safe issues
Install ECC (requires Claude Code CLI v2.1.0+):
/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@ecc
One honest note: the core is MIT and free, alongside a paid Pro tier ($19/seat/mo).
4. Microsoft Agent Governance Toolkit - Destructive Commands Made Impossible
3.6K stars. MIT. From Microsoft.
Prompt-level safety is a polite request to a stochastic system - cited research reports up to a 100% attack success rate on frontier models. AGT intercepts every tool call, message, and delegation in deterministic code before it reaches the wire. Denied actions aren''t unlikely - they''re structurally impossible.
Govern any tool in two lines against a YAML policy:
from agentmesh.governance import govern
safe_tool = govern(my_tool, policy="policy.yaml")
A policy that blocks destructive database operations:
rules:
- name: block-destructive
condition: "action.type in [''drop'', ''delete'', ''truncate'']"
action: deny
What happens when the agent tries anyway:
>>> safe_tool(action="drop", table="users")
GovernanceDenied: Action denied by policy rule ''block-destructive''
Claims coverage of the OWASP Agentic Top 10 and maps to NIST AI RMF, EU AI Act, and SOC 2. Governs Claude Code, Copilot CLI, OpenCode, LangGraph, CrewAI, AutoGen, and more.
pip install agent-governance-toolkit[full]
/plugin marketplace add microsoft/agent-governance-toolkit
/plugin install agt-governance@agent-governance-toolkit
In Public Preview - Microsoft-signed releases, may have breaking changes before GA.