Glossary

Short definitions of the terms used across Liyagent: agents, bindings, guardrails, breakers, A2A, Cedar policies and more.

TermMeaning
A2AAgent-to-agent protocol. Liyagent can publish an agent card and call remote agents.
Binding (model)The provider, model and fallback lists an agent calls.
Binding (guardrail)Which guardrail set applies to which agent.
BreakerPer-backend circuit breaker: closed, open or half-open.
CedarThe policy language Liyagent uses for authorization.
DeflectionAnswering a request without a ticket reaching a person.
Dead modelA model skipped until its reset time after the provider says it is gone or out of quota.
EngineHow a decision is made: rules, rules plus model, or model.
Evidence bundleA signed export of the hash-chained audit log that anyone can verify offline with the instance's published audit signing key.
Known issueAn open ticket that later duplicate reports are folded into.
LiyaLiyagent's agent family, and the agent people talk to.
Managed agentAn agent Liyagent runs: instructions, tools, sub-agents and limits in the console.
MCPModel Context Protocol, how agents get tools.
Run asWhose permissions a managed agent's tool calls use: the agent's own, or the invoker's as well.
Self-managed agentAn agent running elsewhere that calls the gateway with its own credential.
Semantic cacheReuse of a verified earlier answer for a question that means the same thing.
Sensitivity labelpublic, internal, confidential or restricted. Limits where an item may be retrieved.
Sub-agentA specialist a managed agent can delegate to. One defined on the parent spends the parent's budget.
TaxonomyA tenant's categories, default IT or learned from its tickets.
TenantA separated slice of one instance with its own data and settings.
TriggerAn inbound channel bound to one agent.
Write postureWhether an agent's write tool call on an MCP server is allowed, held for approval or denied.