Completing the CI/CD Pipeline for AI Agents: How 3 New Skills Filled Critical Gaps

#mcp #ai #cicd

Completing the CI/CD Pipeline for AI Agents: How 3 New Skills Filled Critical Gaps

The Problem: A Broken Pipeline

In our previous articles, we discussed Lianzhu's five-stage CI/CD framework for AI agents. But there was a gap. Three critical positions in the pipeline were empty.

The Five Stages: Before and After

Stage	Skill	Previous Status	Current Status
Route Definition	QuanYan·Source Seeking	Live	Live
Acceptance Criteria	SunMao·Acceptance Building	Missing	Now Live
Quality Assurance	CuiLu·Skill Forging	Live	Live
Multi-Role Testing	ShaLuang·Multi-Role Trial	Missing	Now Live
Signal Feedback	DuoSheng·Public Signal Indexing	Missing	Now Live

The pipeline was broken at the acceptance criteria, testing, and feedback stages. When running an AI agent task with Lianzhu, you could define requirements, perform QA checks—but then you had no formal way to validate acceptance criteria, test with different user perspectives, or collect real-world signal feedback.

SunMao: Define Acceptance Before Execution

The gap it fills: Acceptance criteria is always added after the fact.

You've probably experienced this scenario: An agent completes a task. You glance at it and think "good enough"—then it ships. Later, problems emerge. You ask: "What were the acceptance criteria?" The answer: none.

SunMao changes this workflow:

Define acceptance conditions first, converting requirements into executable checks
Lock down success and failure paths before execution begins
Include file manifests and hashes to verify deliverables haven't been silently modified mid-execution

The hash verification is particularly important. Since AI agents can modify their own work during execution, SunMao ensures that acceptance validation catches unauthorized changes before sign-off.

ShaLuang: Stress-Test Before Launch

The gap it fills: Internal testing can't substitute for real user perspectives.

The Lianzhu pipeline previously terminated at the QA stage. The three QA gates catch logical errors—but they don't catch "seems fine to me but users hate it" scenarios.

ShaLuang introduces multi-perspective testing:

Pre-defined seven user archetypes (varying experience levels, different use cases)
Each archetype surfaces different blind spots:
- New users get stuck on terminology
- Power users hit edge cases in defaults
- Task-focused users skip optional-but-useful steps
Clear separation between simulated feedback and real user data

This distinction is critical. Many agent testing frameworks conflate simulation results with actual user feedback, leading to decisions built on false data. ShaLuang maintains this boundary: simulation exposes obvious issues; real user feedback requires separate collection.

DuoSheng: Listen to Public Signals, Not Internal Hype

The gap it fills: No systematic way to collect post-launch feedback.

The first four Lianzhu stages complete entirely internally. But what happens after launch? Your Skill goes live, users interact with it, discussions happen online—how do you systematically collect these signals and feed them back into development?

DuoSheng addresses this:

Convert ShaLuang testing questions into public search queries
Log each discovery with source, timestamp, and credibility score
Feed validated signals back to the development origin point
Policy: Listen only to external echoes, not internal celebration

No internal testing data. Only public sources. Every signal logged with provenance, not accepted wholesale.

This has practical value. Ship a Skill or feature, then wonder what users think. A full manual search cycle takes hours. DuoSheng standardizes this workflow.

The Complete Pipeline Now

With all three Skills live, the Lianzhu pipeline is now complete:

QuanYan (Define Route)
  → SunMao (Define Acceptance)
    → CuiLu (Execute Quality Checks)
      → ShaLuang (Multi-Role Testing)
        → DuoSheng (Public Signal Feedback)

This mirrors CI/CD best practices: each stage has clear inputs and outputs. When something fails, you know exactly which stage to debug.

How to Use

All three new Skills are available through MCP. Add https://mcp.tancoai.com/mcp to your MCP configuration and call discover_skills to see the full toolkit.

Use individually: Call each Skill by name for specific needs.

Use as a chain: Call Lianzhu, which orchestrates all five stages in sequence automatically.

New users get 5 free complete-task credits—perfect for experiencing the full five-stage pipeline with a real task.

Explore the complete Lianzhu toolkit: