AI model, agent and harness: what is the difference?

A model generates a response from its input. An agent uses a model to decide what to do next, often calling tools along the way. A harness provides the environment that runs those decisions, manages state and enforces boundaries. These are different parts of one working system, not interchangeable product names.

Use a simple purchasing example

Imagine a team preparing a supplier comparison. The model can extract terms from a document. The agent can decide to retrieve a missing quote before comparing options. The harness controls which supplier folder is accessible, records the retrieved versions and prevents an unapproved purchase.

Changing the model may improve extraction. It will not necessarily repair a missing permission check or a broken folder connection. That is why diagnosing every failure as a model problem wastes time.

Ask where the responsibility lives

• Incorrect reasoning: inspect the task, examples and model behavior.
• Wrong document: inspect retrieval, source labels and access filters.
• Lost progress: inspect the saved task state and restart behavior.
• Unapproved action: inspect the execution policy, not just the prompt.
• False completion: inspect the evidence required after a tool call.

Some failures cross several layers. Save the input, selected sources and action results together so the team can reconstruct the sequence instead of guessing from the final sentence.

Compare proposals by outcome

Ask each provider to demonstrate the same bounded task with the same inputs. Include a normal case, a missing document and a rejected approval. Record whether the result is correct, whether permissions hold and whether a teammate can take over.

Do not compare only model names or the number of agents. A simple design that completes your process can be a better fit than a complicated design with impressive labels. Maintenance ownership also belongs in the comparison: who fixes the workflow after a tool changes?

A useful one-page inventory

Write the model, agent responsibility and harness owner in separate fields. Add the connected tools and the person accountable for the business outcome. If nobody owns recovery, the inventory has revealed a real operational gap before it becomes a customer problem.

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