AI Universe · The AI application (agent harness)

The AI application (agent harness) zoom 2

What surrounds the model call? 7 concept buckets. Each one is a job you can learn once; underneath it, who does that job on each platform.

Concept

Control loop

Ask the model, run the tools it asks for, go round again, stop safely.

steps and budgetsstop rulesworkflow vs agent

Physical: who does this job, where

Open sourceplain Python, LangGraph, Pydantic AI, OpenAI Agents SDK, CrewAI
AzureMicrosoft Agent Framework, Foundry Agent Service
AWSBedrock Agents, Strands, AgentCore runtime
GCPAgent Development Kit, Vertex AI Agent Engine
VendorsClaude Agent SDK

Concept

Context assembly

Building exactly what the model reads on every call, within a budget.

system prompthistory and trimmingfencing untrusted text

Physical: who does this job, where

Open sourceyour own code, LlamaIndex
AzureFoundry prompt flow
AWSBedrock prompt management
GCPVertex prompt management
Vendorsprovider context tools

Concept

Tools and protocols

Letting the model act in the world, safely and typed.

tool schemasMCPagent-to-agent (A2A)

Physical: who does this job, where

Open sourceMCP SDKs, FastMCP, A2A SDK
AzureAPIM MCP, Foundry tools
AWSAgentCore Gateway
GCPVertex AI extensions, ADK tools
VendorsMCP servers from SaaS vendors

Concept

State and memory

Remembering the conversation and facts, across turns and sessions.

conversationlong-term memorycheckpoints

Physical: who does this job, where

Open sourcePostgreSQL, Redis, Mem0, Letta
AzureCosmos DB, Azure PostgreSQL
AWSAgentCore Memory, DynamoDB
GCPAgent Engine memory, Firestore
VendorsZep

Concept

Guardrails and policy

Checks before and after every step: what may go in, come out, and run.

input checksoutput checkstool permissionconsent

Physical: who does this job, where

Open sourceyour own rules, Guardrails AI, NeMo Guardrails, Llama Guard
AzureAI Content Safety, Prompt Shields
AWSBedrock Guardrails
GCPModel Armor
VendorsLakera

Concept

Durable work and people

Waiting hours for a person, surviving restarts, never paying twice.

durable workflowsapprovalshand-offsidempotency

Physical: who does this job, where

Open sourceDBOS, Temporal, Restate
AzureDurable Task Scheduler, Logic Apps
AWSStep Functions
GCPWorkflows
VendorsInngest

Concept

Identity

Who is asking, and whose authority the agent acts with.

user sign-inagent identityon-behalf-of tokens

Physical: who does this job, where

Open sourceKeycloak, PyJWT
AzureEntra ID, managed identity
AWSCognito, IAM, AgentCore Identity
GCPIdentity Platform, IAM
VendorsOkta, Auth0

Zoom in further

← Model access · Experience →

Compute & infrastructureDataModelModel accessThe AI application (agent harness)ExperienceAssurance & operations