Last reviewed: August 2026
If you are new to AI, read Getting Started with AI and Prompt Engineering first. This document focuses on the concepts, architecture, and vendor platforms of autonomous agents that build on top of them. For organizational adoption strategy, see the Agent Adoption Guide .
Traditional LLMs follow a single prompt → single response pattern. AI agents have an autonomous execution loop : given a goal, they plan, call tools, verify results, and retry if needed.
Aspect
LLM Prompting
AI Agent
Execution
Single request-response
Multi-step loop (observe → think → act → repeat)
External integration
Limited
Tool calling (APIs, DBs, filesystem)
Autonomy
User directs each step
Given a goal, decomposes and executes independently
When agents are unnecessary: Simple Q&A, well-defined pipelines (Step Functions, etc.), real-time response requirements.
Type
Audience
Examples
Characteristics
Desktop Agent (work)All employees
Claude Cowork, Amazon Quick, ChatGPT Work, M365 Copilot, Gemini
Local file/app access, Computer Use, MCP connectors
Coding Agent (dev)Engineering
Kiro, Claude Code, Codex, Grok Build, Copilot, Antigravity, OpenCode
Terminal/IDE/Git, code gen/edit/test/PR
Autonomous Ops Agent DevOps/Security/FinOps
AWS DevOps/Security/FinOps Agent, Security Copilot, Google SecOps Agents
Hours-to-days autonomous execution, no constant human supervision
LLM chat was confined to the browser. Desktop Agents break this limit with local file access, OS manipulation (Computer Use), external tool connection (MCP), and long-running autonomous execution.
Aspect
Self-hosted (OpenClaw, Hermes, etc.)
Managed (Claude Cowork, Quick, Copilot)
Deployment
User installs
IT deploys via MDM/SSO
Model
Local / personal API key
Vendor-hosted (frontier models)
Data control
Local control (hard to enforce org policy)
DLP, connector allowlists, audit logs
Advantage
Privacy, customization
Governance, frontier models, enterprise tool integration
Claude Cowork status (2026.08): macOS/Windows GA (April) → web, iOS, Android + cloud remote sessions (July). Chrome side panel integration, cross-device session continuity.
Domain
AWS
Microsoft
Google Cloud
Security
Security Agent (GA)
Security Copilot Agents (GA)
Security Operations Agents (Preview)
DevOps/SRE
DevOps Agent (GA)
Azure Copilot
—
FinOps
FinOps Agent (Preview)
Azure Copilot Cost Optimization
—
Coding
Kiro (IDE/CLI/Web)
GitHub Copilot
Antigravity
Pattern
Description
When to Use
ReAct Alternates reasoning and action
Single agent, simple tool calls
Plan-and-Execute Full plan then sequential execution
Complex multi-step tasks
Multi-Agent Specialized agents collaborate by role
Large workflows, domain separation
Human-in-the-Loop Human approval before risky actions
Production, high-risk operations
Protocol
Role
Key Points
MCP Agent → Tools/Data
2026-07-28 spec : stateless core, Extensions framework, Tasks, MCP Apps. 400M+ monthly SDK downloads
A2A Agent → Agent (cross-vendor)
v1.0 (March 2026 GA). Multi-protocol bindings, signed Agent Cards, 150+ participating orgs
ACP Agent → Agent (internal peers)
REST-native, no SDK required
All three protocols are under AAIF (Linux Foundation) governance.
The largest revision of MCP since launch. Key changes:
Stateless core — Protocol-level sessions (Mcp-Session-Id) and initialize handshake removed. Serverless/edge deployment now possibleExtensions framework — Reverse-DNS identifiers with independent versioning. Tasks and MCP Apps graduated as official ExtensionsTasks — Standard lifecycle for async long-running operationsMCP Apps — Server-rendered interactive UI sandboxed at the hostAuthorization hardening — OAuth 2.1-based authorization improvementsFormal deprecation policy — Roots, Sampling, and Logging marked deprecated
Already supported by AgentCore Gateway and Claude products.
Vendor
Platform
Strengths
AWS
Bedrock AgentCore Framework-agnostic, Harness, Memory, Gateway, MCP
Azure
Microsoft Foundry Agents Responses API, MCP, Agent 365 governance
Google
Gemini Enterprise Agent Platform ADK (open-source), A2A native, Agent Runtime
OCI
OCI Enterprise AI Agents RAG agents, Oracle DB integration, AI Guardrails
Framework
Characteristics
LangGraph State machine-based multi-agent
CrewAI Role-based collaboration
Strands Agents AWS open-source, model-agnostic
AutoGen Microsoft, conversational multi-agent
Product
Provider
Characteristics
Kiro AWS
Spec-driven, Hooks, IDE/CLI/Web
Claude Code Anthropic
Agent Teams, 29 hooks, plugins
Codex OpenAI
Parallel agents, Computer Use
Grok Build SpaceXAI
8 parallel sub-agents, Git worktree isolation
GitHub Copilot Microsoft
Agent Mode, Agent Merge, Cloud Sessions
Antigravity Google
Agent-first IDE, Managed Agents
OpenCode Anomaly
Open-source, model-agnostic
Area
Details
Cost Loop execution consumes 10–100× tokens. Per-task budgets, loop limits, model tiering required
Evaluation Task success rate, tool selection accuracy, hallucination rate
Observability OpenTelemetry-based tracing. Agent-specific metrics in LLMOps
Security Prompt injection, privilege escalation, data exfiltration, infinite loops. Details in AI Security
Risk
Mitigation
Long-running session cost explosion
Session budgets, auto-termination
Cross-app injection
Connector allowlists, input sanitization
Autonomous agent drift
Checkpoints, kill switch, diff review
Shadow AI
Provide equivalent experience via official Desktop Agent