Skip to content
AI Work Orchestration

Agent Runtime vs Orchestration Layer vs Framework

The three layers of production AI agent infrastructure. Frameworks build agents. Runtimes run them. The orchestration layer turns them into a team. You need all three.

Curate-Me is the work orchestration platform that works with any agent runtime and any framework. Add routing, cost control, PII scanning, and human approvals in two minutes.


The agent infrastructure taxonomy

Three distinct layers, three distinct jobs. Most teams conflate them -- and end up with unmanaged agents in production.

 
Framework
Runtime
Orchestration
What it doesProvides building blocks -- prompts, chains, tool use, memoryRuns the agent -- loops, error recovery, tool executionOrchestrates the work -- routing, cost limits, approvals, audit trail
AnalogyThe blueprintThe workerThe control room
ExamplesLangChain, CrewAI, Vercel AI SDK, Semantic KernelClaude Code, OpenClaw, Codex, OpenCode, DevinCurate-Me, Portkey, Kong AI Gateway, LiteLLM
ScopeSingle agent logicAgent lifecycle and execution environmentCross-agent orchestration and policy enforcement
Runs whereInside your app processIn a container or sandboxAs the orchestration layer between agents and LLM providers
Failure modeBad prompt, wrong tool choiceInfinite loop, resource exhaustionBudget blown, data leaked, model misuse

Taxonomy based on LangChain, Anthropic, and Thoughtworks definitions (2026).

You need all three layers

A framework without a runtime is a script. A runtime without an orchestration layer is an unmanaged autonomous system. Production agents require all three.

Framework
LangChain, CrewAI, etc.
Runtime
Claude Code, OpenClaw, Codex, etc.
Orchestration
Curate-Me
Provider
OpenAI, Anthropic, etc.

Framework: LangChain, CrewAI, etc.

Provides the primitives: prompt templates, tool definitions, chain composition, and memory management. The framework gives you building blocks, but it does not run or govern anything.

Runtime: Claude Code, OpenClaw, Codex, etc.

Runs the agent in a loop: receives a task, calls the LLM, executes tools, handles errors, and iterates until done. The runtime makes the agent work, but it does not control what the agent is allowed to do.

Orchestration: Curate-Me

The control room for your AI team. Every API call passes through an orchestration chain: routing, cost caps, PII scanning, model selection, and human approval gates. Curate-Me turns individual agents into a managed workforce.

The key insight: Most teams start with a framework and add a runtime when they need autonomous execution. The orchestration layer is what they forget -- until an agent blows through a budget, leaks PII, or calls a model it should not have access to. By then, the damage is done.


Where Curate-Me fits

One control room for every AI agent your team uses. Curate-Me is the work orchestration layer -- it works with any agent runtime and any framework.

Zero code changes. One base URL swap.

# Before (direct to provider):
OPENAI_BASE_URL=https://api.openai.com/v1

# After (through Curate-Me):
OPENAI_BASE_URL=https://api.curate-me.ai/v1/openai
X-CM-API-Key: cm_sk_xxx

Cost Governance

Per-request cost limits and daily budgets. Real-time spend tracking in Redis with MongoDB audit trail. No surprise bills.

PII and Secrets Scanning

Regex-based detection for API keys, credit card numbers, SSNs, and personal data. Requests are blocked or redacted before they reach any provider.

Security Scanner

Prompt injection, jailbreak, and data exfiltration detection. Six-stage safety pipeline that short-circuits on first denial.

Human-in-the-Loop Approvals

High-cost or sensitive operations are held in an approval queue. Reviewers see full context and approve or deny with one click.

Immutable Audit Trail

Every request, governance decision, and runner action is logged to an append-only audit log. Time-travel debugging lets you replay any execution.

Managed Runners

Sandbox containers for Claude Code, OpenClaw, and custom agents with lifecycle management, network phase separation, and 3-tier tool profiles.

Works with every agent runtime and framework

Claude CodeAnthropic's coding agent
OpenClawOpen-source personal AI assistant
LangChain / LangGraphChain and graph-based agents
CrewAIMulti-agent orchestration
Vercel AI SDKStreaming-first AI toolkit
Custom agentsAny code that calls an LLM API

How agent infrastructure evolved

The concept of layered agent infrastructure emerged across several organizations in early 2026. Here is the timeline.

Jan 2026Philipp SchmidHugging Face

Describes the harness as "the OS for AI agents" -- the runtime that manages the agent loop, tool execution, and context window.

Mar 2026AnthropicAnthropic

Publishes guidance on "harness design for long-running applications" -- the outer loop that manages agent lifecycle, error recovery, and human oversight.

Mar 2026LangChainLangChain

Publishes "The Anatomy of an Agent Harness" -- distinguishing the harness (runtime loop) from the framework (building blocks) and the orchestrator (multi-agent coordination).

Apr 2026Martin FowlerThoughtworks

Coins "harness engineering" as a discipline -- the practice of building guides (prompt constraints) and sensors (observability, cost tracking) around autonomous agents.

Apr 2026SalesforceSalesforce

Publishes their "agent harness" definition page, positioning the harness as the control surface between human intent and agent execution.

The emerging consensus: An agent runtime is the execution wrapper around an AI agent -- the loop that manages execution, error recovery, and tool use. It is distinct from the framework (building blocks) and the orchestration layer (routing, governance, audit). Production agents need both “guides” (constraints) and “sensors” (observability). The orchestration layer provides both.


Add orchestration to any agent in 2 minutes

One platform for every AI agent your team uses. Route work to the right model, enforce budgets, scan for PII, and keep the full audit trail.

Free tier -- no credit card
26 production LLM providers
Works with any agent