Skip to content
be-human.ai
← Back to the journal

Harness fundamentals

Harness Versus Prompt Files: What Superintelligence Buyers Should Ask

A prompt file holds the instructions. The harness holds tools, memory, permissions, checks and recovery. Learn what to ask about each before you buy.

Be Human typographic cover with the title: Harness Versus Prompt Files: What Superintelligence Buyers Should Ask
Conceptual illustration created with AI · Be Human

A prompt file is the text that tells a model how to behave on a task: its role, tone, rules, output format and examples. The harness is everything around that file that runs the work, including model choice, tools, memory, permissions, validation, logging and recovery after a failure. When you evaluate an agent, the prompt is usually the smallest part of what you are buying, and most of the reliability and risk sits in the harness.

Key takeaways

  • A prompt file sets instructions for a task. The harness is the operating layer that runs that task against real systems.
  • Judge a prompt by its behavior under test, not by its length or wording.
  • Ask who controls tools, permissions, memory and approval points, because those decide what the agent can actually do.
  • Require logs, stop conditions and a rollback path before any agent touches a live system.
  • Treat a model change as a harness change that needs review, not as a quick edit to the text.

Why the prompt file gets all the attention

Most early agent projects begin with a document. Someone writes a long block of instructions, pastes it into a configuration field, and the first demonstration looks convincing. The prompt is visible, easy to edit and easy to explain to a manager, so it becomes the center of the conversation. That focus is understandable, but it pushes buyers toward the wrong first question. A well-written prompt can still produce a bad outcome when the agent has the wrong tools, reads outdated records, retries an action twice or keeps working after it has lost track of the goal.

What a prompt file controls and what it cannot

A prompt file shapes how the model interprets a request. It can set the role, the order of steps, the format of the answer and the kind of judgment the agent should apply. It deserves careful writing and regular review. But the file enforces nothing by itself. A sentence that says "never issue a credit above the approved limit" is a request the model will usually follow. It is not a permission. The real limit has to live in the tool's permissions, in code or in an approval step the model cannot bypass.

Prompt files also cannot separate instructions from content. An email, a web page or a customer note can contain text that tries to override the agent's rules. The OWASP prompt injection guidance treats this as a design problem that needs controls outside the prompt, such as limiting what untrusted content can trigger. The defense is a property of the surrounding system, not of a better sentence in the file.

What the harness is made of

The harness is the set of parts that operate during a run. In most business deployments it includes:

  • Model selection and fallback, which decides which model handles each step and what happens when it is slow, unavailable or changed.
  • Tools and their input rules, which define the actions the agent may take and the fields each action accepts.
  • Memory and context assembly, which decides which records the agent reads, how current they are and when they expire.
  • Permissions and approval points, which decide what runs alone and what waits for a person.
  • Validation, logging, stop conditions and recovery, which show what happened and allow interrupted work to resume safely.

Anthropic's engineering notes on long-running agents describe a similar split: much of what makes long work reliable is the structure around the model, including how progress is recorded and how a new session picks up the thread. For a plain definition of the layer, see what is an AI harness, and for a checklist you can use with any vendor, see AI harness requirements.

Which questions to ask about the prompt file

When a vendor or internal team shows you a prompt file, the goal is to learn whether it is maintained, testable and bounded. Ask these questions:

  • Who wrote it, when was it last changed, and what changed?
  • Does each important rule have a test case that shows the agent follows it?
  • Is the file short enough for an operations lead to read in one sitting, with long procedures moved into separately versioned documents or skills?
  • Does it state what the agent must refuse or escalate, not only what it should try to do?
  • Does it tell the agent to treat customer messages, attachments and web content as data rather than as instructions?
  • Which parts would need rewriting if the underlying model were replaced?

A good answer to these questions is specific. It names a file, a test and an owner. A vague answer such as "we tuned it extensively" tells you that nobody can show the evidence.

Which questions to ask about the harness

The harness questions are where most of the business risk becomes visible. Ask:

  • Which tools can the agent call, and which of them write to a live system such as an accounting ledger, a calendar or a customer inbox?
  • Who approves a new tool, a wider permission or a new data source, and where is that decision recorded?
  • Which model handles each task, and what happens when that model fails, times out or behaves differently after an update?
  • What data does the agent read before deciding, how current is it, and what does it do when a record is missing?
  • If a run is interrupted, can it resume from a verified point, or does it start over and risk repeating an action?
  • Where are the logs, and can a person reconstruct why the agent took each action on a particular day?
  • How are changes tested before reaching production, and who can roll them back?

For the guardrails that matter most at the start, see AI agent guardrails for business.

An illustrative scenario and a review plan

Illustrative scenario: a regional distributor wants an agent to draft replies to customers asking about order status. The vendor delivers a polished prompt file with a friendly tone and clear escalation language. During review, the team discovers that the agent reads order data from a spreadsheet updated once a day, and that the same connection can issue credit notes with no approval step. The prompt is adequate. The harness is not, and the review found that before any customer saw a reply.

A practical way to review a setup like this follows a short sequence:

  • Step 1: Write the single task and the result that counts as success, in terms someone can check.
  • Step 2: List every tool and data source the agent may touch, and mark each one that can write.
  • Step 3: Remove write access for the first release, or route every write through an approval step.
  • Step 4: Test the prompt against real cases, including messages that try to override its rules.
  • Step 5: Confirm that logs, stop conditions and resume behavior exist before go-live.
  • Step 6: Record each change in a release note so the next review knows what moved and why.

How Be Human handles the harness around the prompt

Be Human designs and operates the harness as part of the service, not only the prompt. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. The prompt is then written against that definition and tested inside it. After go-live, managed operations cover monitoring, recurring evaluations, permission review, integration maintenance, incident response and controlled improvements, so that tool scopes and instructions are reviewed on a schedule rather than drifting unnoticed.

If you are unsure which task should go to an agent first, the free Be Human AI diagnostic can help you find the first task worth handing over.

Frequently asked questions

Is a prompt file the same as a system prompt? A system prompt is one common form of prompt file, supplied to the model at the start of a session. Many agents also load instructions from workspace files, skills or task definitions, so the prompt is really one layer of instructions inside the harness. Whatever name a vendor uses, ask where each instruction lives and who is allowed to change it.

Can a better prompt replace a harness? No. A prompt can describe the behavior you want, but it cannot enforce permissions, validate outputs, resume interrupted work or record what happened. Those jobs belong to software and operating procedures outside the model. A strong prompt inside a weak harness still carries risk that the wording cannot remove.

Do I need to build the harness myself? Not necessarily. Some teams prefer an operator to design and run the harness while they keep ownership of business rules, data and approvals. The right choice depends on how often your tasks change, who will maintain the integrations and how much internal engineering time you can commit to them. The comparison in AI harness: build or buy walks through those tradeoffs.

What should I ask a vendor first? Ask for the prompt file and the harness description side by side. Request a tool list with write access marked, the approval points, a sample of the logs and the plan for changing the underlying model. A vendor who can explain each of these in plain language is easier to evaluate than one who only shows a longer prompt.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →