Agentic

Agentic AI & multi-agent systems

LLM agents that plan and call tools — the hard parts are tool safety, isolation, and cost control.

LangGraphCrewAIRay ServevLLMMCP

Architecture

An orchestrator runs agent loops that call an inference backend and a set of tools. Tools are sandboxed and scoped per tenant; actions are logged and gated.

Sizing

GPU demand is the inference backend; agents multiply requests, so batch and cache aggressively (prefix caching).

Networking

Standard; isolate tool egress and enforce per-tenant network policy.

Storage

Object storage for artifacts/traces; a state store for agent memory.

Scaling

Scale the inference backend on token throughput; rate-limit tool calls to bound cost.

Monitoring

Tokens/sec, tool-call rate and latency, cost per task, guardrail triggers, error rate.

Sample manifest
starting point
apiVersion: apps/v1
kind: Deployment
metadata:
  name: agent-orchestrator
spec:
  replicas: 2
  template:
    spec:
      containers:
        - name: orchestrator
          image: your-registry/agents:latest
          env:
            - { name: LLM_ENDPOINT, value: "http://llm/v1" }
            - { name: TOOL_ALLOWLIST, value: "search,retrieve" }
Recommended GPUs