The Agent is a Process: Why Harness and Loop Define AI's Future

An agent is not an app with an LLM stuffed inside. An agent is a process running in a second-level operating system, where the harness defines the boundaries and the loop manages execution. This thesis — from Stefano Maestri's deep dive on June 15, 2026 — reframes the entire agent architecture conversation. The LLM is the engine. The harness is the chassis. The loop is the transmission. Together, they form a new minimal unit of work that we can move across machines, networks, and clouds — something closer to a pod than a microservice. (Source: artificialcode.substack.com, June 15, 2026)

[ STAT ] The agent is a process running in a second-level operating system — harness defines boundaries, loop manages execution. — Stefano Maestri, 2026

What This Actually Means

Recent industry moves confirm this thesis. OpenAI acquired Ona (formerly Gitpod) to give Codex secure cloud environments and orchestrate persistent tasks. Xiaomi released MiMo Code — an open-source coding harness that holds up over sequences beyond 200 steps with persistent memory. NVIDIA published SkillSpector to analyze agents' skills for vulnerabilities. LangChain now offers hardware-isolated microVMs for each agent. The pattern is clear: the industry is no longer building applications with LLMs inside. It's building on top of the agent as the fundamental unit.

[TOOL: Harness] Defines agent boundaries: tool access, data scope, execution environment, permission policies, and cost limits. Examples: Claude Code dynamic workflows, Omnigent meta-harness, MiMo Code harness.

[TOOL: Loop] Manages execution: plan-act-observe-adapt cycle, state persistence across iterations, termination detection. The loop replaces the human prompter.

Who This Is Built For

For AI engineers and architects: the harness-and-loop model gives you a clean architectural pattern for agent systems. The harness is the contract between the agent and the infrastructure. The loop is the execution engine. Build with this separation and you can swap models, tools, and environments independently.

For platform engineering teams building agent infrastructure: your goal is to provide the harness layer — standardized execution environments, tool registries, policy enforcement, and observability. Individual teams build loops within your harness.

For enterprise architects evaluating agent platforms: evaluate platforms on their harness capabilities — not just model support. A platform with good harness features (isolation, policies, cost controls) will outperform one with better models but worse harnesses.

Key Architectural Insights

Harness Is the Contract: The harness defines what the agent can do, what data it can access, what resources it can consume, and when it must stop. A good harness prevents the agent from exceeding its boundaries without blocking legitimate work. A bad harness either cages the agent (too restrictive) or leaves the door open (too permissive).
Loop Is the Execution Engine: The loop manages the plan-act-observe-adapt cycle. It tracks state across iterations, detects goal completion or impossibility, and manages resource consumption. The loop is what makes an agent autonomous — without the loop, you have a single-turn LLM call.
The Agent Is a Deployable Unit: With a well-defined harness and loop, the agent becomes a package that can be deployed on any infrastructure that supports the harness interface. This is the direction the industry is heading — agents as portable, manageable units of work.

The Numbers

▸ OpenAI acquired Ona (Gitpod) to give Codex secure cloud environments for persistent agent tasks ▸ MiMo Code: open-source coding harness, 200+ step sequences, persistent memory via subagents ▸ NVIDIA SkillSpector: analyzes agent skill plugins for vulnerabilities before installation ▸ LangChain microVMs: hardware-isolated agent execution environments ▸ Agent as pod: portable across machines, networks, and clouds (Source: Multiple, 2026)

What This Cannot Do

The harness-and-loop model is an architectural framework, not a product. No single vendor implements it completely yet.
The "second-level operating system" concept assumes standardized infrastructure that doesn't fully exist yet. Current implementations are fragmented across vendors.
Portable agents require standardized harness interfaces. The industry is moving in this direction (MCP, A2A, OpenShell) but hasn't converged on a single standard.

Start in 10 Minutes

(2 min) Read the full essay: artificialcode.substack.com (June 15, 2026)
(3 min) Map your current agent setup to harness vs. loop — what defines boundaries? What manages execution?
(5 min) Identify one improvement: is your harness too restrictive or too permissive? Does your loop handle failures gracefully?

Frequently Asked Questions

Q: Why is the harness more important than the model? A: The model determines what the agent can understand. The harness determines what the agent can do. A mediocre model with a great harness (good tools, clear boundaries, cost controls) outperforms a great model with a poor harness. (Source: Stefano Maestri, 2026)

Q: How does this change how I build agents? A: Build with the harness-loop separation from day one. Define the harness first (what tools, what permissions, what costs), then build the loop (what goal, what success criteria, what termination rules). The model is a configuration choice within this architecture.

Q: Is this relevant for simple agents? A: The harness-loop model scales down as well. Even a simple email classification agent benefits from clear boundaries (only read email subject lines, not body content) and a defined loop (classify → if confidence < 0.8, escalate → log result).