Core Announcement: Agent Loop Open-Source Controller Skill Release
The open-source Skill project Agent Loop has released its stable version v1.5.8 (MIT license), positioned as a controller-type Skill rather than a standalone Agent or platform. Its core purpose is addressing the closed-loop delivery gap in real-world projects involving current AI coding agents.
- Release date: Stable v1.5.8 available, source code open under MIT license
- Supported agents: Claude Code, Codex, Kimi Code CLI, OpenCode, and other mainstream CLI Code Agents
- Installation: One-command via
npx skills add, plus Git clone path for non-npx environments - Working mode: Designed for “one human + one CLI Agent” workflow without disrupting existing habits
Root Problem: AI Coding Is Not Short of “Writing Code”, But of “Closed Loops”
Current AI coding agents (Claude Code, Kimi Code CLI, etc.) have mature code-generation capabilities, yet industry traading reveals five pain points in real projects:
- No workflow: Starting implementation before requirements are clarified, lacking verificationpost implementation;
- No memory: Forgetting project context across conversations, repeating past decisions and mistakes;
- No boundaries: Remaining silent when asking is needed, but executing dangerous production/release actions autonomously;
- No closure: Treating “code done” as “task done”, missing testing, review, and recovery;
- No evolution: Absent lifecycle management, leading to drift between requirements, code, and documentation.
Unexpected finding: Agent Loop isn’t a new AI model or framework—it’s a standard Skill embedded into users’ existing tools. Developers gain closed-loop capabilities without switching tools or workflows, significantly reducing adoption barrier.
Core Design: Human-Directed, Agent-Owned, Evidence-Verified
Agent Loop follows the “Human-directed · Agent-owned · Evidence-verified” philosophy, with the main loop:
Human Goal → Controller Phase Detection → Product Definition → Design Readiness (ADR) → Graduated Delivery → Verification & Review → Memory Closure
Its five core capabilities:
Project Understanding & Takeover: Message Intent Guard categorizes user input (chit-chat/requirement/maintenance/feature/bug), preventing unnecessary workspace creation; Project Entry Scan produces evidence-backed project documentation (
.agent-loop/onboarding-db/);Adaptive Product Definition: Selects brief or standard depth based on uncertainty; outputs Product Model with角色、权限、命令、事件、流程、状态、异常;
Minimal Safe Workflow: Graduated delivery by risk level; supports full-workspace Git quick-commit; v1.5.8 introduces full-test authorization—proposed test matrix MUST specify commands, branches, environments, costs, and conclusions, with single authorization executing once only;
Verification & Closure Chain: Enforces Execute → Verify → Review → Drift Check → Memory Update → Completion Check; initial features and explicit bugs require real RED/GREEN test evidence;
Persistent Project Memory:
.agent-loop/stores stable facts, pending work, recovery points; single command “continue previous Agent Loop task” restores from safest point across sessions/machines.
Human-Agent Responsibility Boundary
Agent Loop clearly delineates authorities:
| Agent Authority | Human-Reserved Authority |
|---|---|
| Query evidence first, determine phases, plan depth | Define goals, scope, product semantics |
| Implement, test, verify, review, drift fixes | Major product/technical decisions |
| Continue until verification or Human Gate | Production/monetary/keys/destructive actions |
| Recommend next steps | Branch changes, commits, push, PR, merge, tag, release |
Critical principle: Approving one gate never equals approving another—tag, push, release, publish are all independent human decisions.
Adoption Recommendations
- Ideal users: Solo developers or small teams relying on CLI agents, frustrated by “code done = task done” mentality;
- Wait briefly: Teams already using mature collaboration platforms (Jira + CI + PR workflow) should assess integration costs;
- Try experimentally: Rapid prototyping, personal learning projects, teaching demos—scenarios where lightweight workflow failsafe is most needed.
Final Note
AI coding is shifting from “capability competition” to “engineering reliability.” Agent Loop compresses industrial-grade quality assurance into reusable Skills for single-developer workflows—not replacing agents, but teaching them when to stop, when to proceed, and how to verify.