A model alone is not an agent — why the harness matters

TL;DR. An agent is not just a clever model. It is a model plus tools, memory, permissions, sandboxes, and the loop that decides the next step. Vendors now ship that scaffolding as a product — the harness. Ask what is governed outside the weights.

The demo that hides the scaffolding

A polished agent demo usually shows:

• A fluent answer.
• A tool call that “just works.”
• A long task that somehow resumes after lunch.

What it rarely shows is the boring machinery: session memory, sandbox limits, policy checks, retries, audit logs, and who is allowed to click “run.” That machinery is the harness — and it is often where safety and usefulness actually live.

A kitchen picture

Piece	Kitchen analogy
Model	The cook’s skill
Tools	Knives, oven, delivery apps
Memory	The order ticket and recipe card
Sandbox / policy	Health rules and “no bare hands in the fryer”
Harness	The whole kitchen layout that makes a rush service possible

Hiring a skilled cook (a strong model) without a kitchen still leaves you with chaos at dinner service.

What to ask vendors (and your own team)

Clarify	Good enough answer sounds like…
Tools	“Here is the allow-list — and what happens on a denied call.”
Memory	“Session vs long-term; what is stored; how it is wiped.”
Sandbox	“Code / browser / network stay inside these walls.”
Human gates	“These steps require approval; these can auto-run.”
Eval	“We score multi-step trajectories on your jobs, not only chat quizzes.”
Pinning	“We pin model and harness version — silent upgrades re-trigger eval.”

Soft OrchAI bridge (one short beat)

OrchAI Agents is the suite’s idea of a governed catalog and control plane — reusable agents under policy, not unbound autonomy. Neighbouring products help: Library for which model routes, Eval for whether the loop still works after a change, Portal for the operating surface. This post is literacy about the industry shape; it is not a claim that any named third-party harness equals OrchAI.

What this post is not

• Not a review or ranking of DeepSeek Harness, Muse, AgentCore, or Gemini Spark.
• Not a mandate to adopt any open-source harness.
• Not a roadmap change.

See also

• Strategy notes: DeepSeek Harness / V4-Pro · Muse harness co-train · NOOA / Open Secure AI · Gemini 3.7 Flash
• Sibling literacy: Same API name ≠ same model · What governance AI means
• Series home: 科普 direction