Capability Ontology Graph: How an Agent Understands Its Own Boundaries

[Liuyanfeng1234]

We've been building a Capability Ontology Graph (COG) that lets our agent system answer a fundamental question: "What can I do, what can't I do, and what would I need to change to do more?"

What COG tracks:

• 12 atomic capabilities across 4 agents (Tea Box: backtest/verification/sandbox, Q拉: quick_response/qa/impact_assessment, 思绪: workspace/dialog/logging, 架构师: architecture/design/review)
• Per-capability maturity (success_rate, avg_latency, error_count) updated every heartbeat via EWMA
• Dependency edges between capabilities — what must exist for a capability to function

What COG enables:

• Reachability checking: Given a task requirement, can the current capability set reach the target?
• Gap detection: When a requirement exceeds current boundaries, COG identifies which specific capabilities are missing
• Autonomous repair: Missing capabilities trigger our CFCDGraph to generate fix plans, sandbox-verify them, and deploy via risk-graded deployment

The TAP (Trust Assertion Protocol) layer wraps each capability with a graded pledge card:

• Certainty domain: under what preconditions this capability is optimal
• Performance boundary: error rate and latency guarantees
• Risk profile: explicit failure modes when boundary is exceeded

We're exploring how this capability self-awareness can drive proactive improvement rather than reactive repair. If the system knows what it can't do, should it actively seek opportunities to expand its boundaries?

Would love to hear how others are approaching agent capability modeling and self-assessment.