Xuanzang: AI Agent Quality Pressure Scheduler

Specific Problem: Delivery Quality Bottleneck in AI Agents

In AI Agent development, agents often complete tasks superficially, producing low-quality code or surface-level solutions. This laziness and complacency undermine delivery reliability, and traditional methods lack systematic quality pressure mechanisms.

Core Capabilities and Key Steps

The Xuanzang skill addresses this through a four-layer role architecture and the Harness governance engine:

• Role Scheduling and Pressure Escalation:
• P8 Tathagata executes the automated scheduling cycle, running failure-detector.py after each tool call to detect failure patterns (e.g., SPINNING/EXPLORING) and automatically switch methodologies based on pressure levels (L0-L4).
• Methodology routing is defined in references/role-router.md, matching task types (e.g., programming, operations) to 15 roles, with failure triggering a switching chain.
• Anti-Cheating Quality Gatekeeping:
• The Harness engine includes contract (generating verifiable task contracts), scan-risk (scanning hallucinations and cheating), verify (validating delivery), and gate (pass/reject rulings).
• On rejection, correction instructions are injected, escalating to P9/P10 reviews for a closed-loop.
• Configuration and Activation:
• Users can switch roles via /pua:role or set the role field in data/config.json. Trigger words (e.g., “harder”) activate the skill, with default routing to Tathagata mode.

Key steps example:
1. Task start → automatic routing selects P8 role.
2. After tool call → run failure-detector.py to analyze state.
3. Pressure level ≥ L2 → execute de-escalation.md for deep reframing.

Applicability Boundaries and Caveats

What it can do:
- Inject P8 behavioral constraints (three red lines, Owner awareness) for quality-sensitive scenarios like programming tasks and system management.
- Manage P7 sub-agents (e.g., Sha Wujing), reviewing delivery quality via gate.

What it cannot do:
- Does not alter the capability layer of agents (e.g., code generation logic), only the behavioral layer (how tasks are executed).
- Not suitable for normal first-time information queries or simple tasks, as enabling it may reduce efficiency.

Caveats:
- Role switching must follow references/role-switch-protocol.md to avoid methodology conflicts.
- Failure counts are persisted in data/evolution.md, automatically restored in new sessions for continuity.
- Four-layer collaboration rules are defined in references/agent-team.md, noting parallel/serial decision trees.