Agents

The Agent plane

Agentic AI as a first-class plane. 24 persistent entities. Typed plans. Branchable. Same governance as humans.

Vadyl's Agent plane is fully native — agents are first-class platform citizens. They read through the access engine, write through the runtime bridge, plan in a typed IR, execute on the seven canonical execution surfaces, bill against project quotas, and are governed by the same envelope as humans. There is no special "agent escape hatch".

What ships

24 persistent entities cover the full agent lifecycle:AgentDefinition, AgentRun, AgentStepRun, AgentReasoningCheckpoint,AgentPlan, AgentPlanStep, AgentMemoryFact, AgentMemoryNamespace,AgentSkill, AgentToolBinding, AgentModelBinding, LlmConnectionBinding,McpExposure, McpClient, TokenAccountingRecord, RecallJournal, and more.

Manifest slot

AgentManifest is the latest branching manifest slot (added in SnapshotManifest v15). Agent definitions, skills, model bindings, and memory namespaces are branchable just like schema and handlers — diff them, sandbox them, propose them, deploy them through the canonical lifecycle.

Capability-aware model routing

Agent authors declare model requirements by capability:

model: llm.require({
  tier: "frontier",
  capabilities: ["tool-use", "vision", "long-context"],
})

ILlmModelRouter resolves through ILlmCapabilityRegistry against the project's LlmConnectionBinding chain with DeprecationFallbackPolicy. Switching from Claude to GPT-5 to a self-hosted model is a binding change — the agent definition stays the same.

Built-in adapters

Anthropic, OpenAI, DeepSeek, Moonshot ship under Vadyl.Connectors.NativeAdapter/ModelInference/<Provider>/V<N>/. Provider SDKs never leak past the adapter boundary; failures are typed ModelInvocationFailureKind (anti-pattern #89 — never Message.Contains).

Hosted vs self-hosted

Hosted LLMs ship as CapabilitySurfaceKind.ModelInferenceAdapter = 8. Self-hosted inference (vLLM, Ollama, GPU pools, on-prem clusters) is realized through CapabilitySurfaceKind.RuntimeSubstrate = 3 as operator-managed inference workloads with the same autoscale, accelerator, resource-class, and vertical resource policy used by project services. Same canonical contract; different substrate.

Run lifecycle

Trigger: API call, scheduled job, event consumer, or human prompt.
Plan: agent generates a typed plan IR, validated through 13 stages.
Approve: optional human-in-the-loop gate based on risk score.
Execute: each plan step dispatches through the canonical pipelines.
Recall: memory facts inform reasoning at each step.
Reconcile: token accounting reconciled, billing emitted, observability flushed.

Sub-agents and authority

Agents can spawn sub-agents — but sub-agent grants only ever narrow, never widen. Parent grants intersected with child override + agent-definition policy + project capability grant authority. The canonical safety guarantee.

Same execution surfaces

Agent runs use the existing seven execution surfaces — durable workflows + triggers — NOT an 8th surface. Pinned by AgentPlaneCoherenceTests.ExecutionSurfaceKind_RemainsSeven.