What it is
A capability registry is a database (typically a small set of tables) that decouples a defined skill or agent capability from the specific model that executes it. The registry tracks capability classes — long-context reasoning, structured JSON output, tool use, vision, code generation — and which models support each class at which performance level. At execution time, a resolver consults the registry plus runtime context (cost ceiling, residency requirements, account constraints, live performance data) and picks a specific model. The same skill record can run on Claude one day and a local Qwen the next, with no skill rewrite. Capability registries are the architectural piece most missing from the agent stack of 2024–2025; they are the substrate behind the "vendor-neutral AI" pattern that emerged in 2026.
Why it matters
Without a capability registry, every "this skill can run on multiple models" claim is implemented through ad-hoc if-statements scattered across the codebase, and inevitably falls behind as new models ship. Skills become vendor configurations again. With a registry, capability is a structural property — auditable, versioned, replayable. Routing decisions are explainable ("we ran skill X on Claude because the resolver picked it given these constraints"). Performance feedback closes the loop: the registry gets smarter about which model serves which capability best, for which workload type, at which cost. This is the difference between "model routing" (a marketing claim) and a real vendor-neutral architecture.
Key components
- Capability classes — reasoning, structured output, tool use, vision, long context, code
- Model support matrix — which models serve which classes at which performance levels
- Resolver integration — runtime selection based on capability + context + policy
- Performance feedback loop — execution outcomes feed back into routing decisions
- Versioning — registry changes are tracked over time as models ship and shift
Related terms
Agent Governance
The policies, controls, and monitoring systems that ensure AI agents operate safely, compliantly, and within business-approved boundaries.
Agent Operations
The discipline of running AI agents in production — capturing what they do, attributing what it costs, evaluating what they produce, and intervening when something goes wrong. The operational layer above agent observability and orchestration.
Agent Infrastructure
The runtime, network, and tooling substrate that AI agents need to execute reliably — sandboxed compute, tool access, memory, gateways to LLM providers, and the orchestration plumbing that connects them. Closer to the metal than agent operations.
LLM Gateway
A unified proxy in front of multiple LLM providers that captures every call, enforces policy, and lets a single application talk to Anthropic, OpenAI, xAI, Gemini, and local models through one interface.
Vendor-Neutral AI
An architecture pattern where AI capabilities — skills, agents, evaluations — are defined separately from the LLM vendor that runs them, so the same capability can execute on Anthropic, OpenAI, xAI, Gemini, or local models without rewriting.