Models, agents, and platforms: what are you actually choosing?

You see an announcement for a new model, an agent demo, and a platform offering dozens of AI tools. They can look interchangeable in a launch feed. They are different things, and confusing them makes comparisons expensive.

Consider one job: turn a folder of support tickets into a weekly report. A model can help interpret the tickets. An agent can organize several steps and use tools. A platform supplies the interface, integrations, runtime, or billing around that work.

1. The model generates or interprets information

A model takes supported inputs and produces supported outputs. A text model might categorize a ticket. A speech model might transcribe a call. A video model might generate a clip from a shot description.

A model name does not tell you which files your application can access, whether your work survives a browser refresh, or whether a report will be emailed. Those behaviors depend on the surrounding product and software.

For the ticket report, a useful model test is narrow: given this ticket and these category definitions, return one category and the supporting sentence. Compare the result with a human-reviewed answer. Keep an unknown category for tickets that do not fit.

2. The workflow connects steps

A workflow defines how the job proceeds. It might load the CSV, remove personal details, classify each ticket, calculate category counts in code, and assemble a draft report.

You do not need an autonomous agent for every workflow. If the steps are predictable, ordinary application code can control their order. A language model can handle the parts that require interpretation while deterministic code performs counting and validation.

For our report, do not ask the model to guess category totals from a long collection of tickets. Calculate totals from the reviewed classification records. Keep the ticket IDs so a reader can inspect the evidence.

3. The agent can choose actions within a task

An agent combines a model with instructions, tools, state, and a loop for making progress. Depending on the implementation, it may decide which source to read next, call a tool, inspect the result, and revise its approach.

The word agent does not guarantee reliable completion. Ask what the system can actually do, what it can remember, when it stops, and which actions require your review.

For the support report, an agent might discover that a file is missing and ask for it. That is different from silently inventing the missing week. It might draft an email containing the report; sending that email is a separate action with different consequences.

4. The platform supplies the working environment

A platform can provide hosting, a user interface, connectors, logs, deployment tools, and account controls. It may expose its own models, models from other publishers, or both.

A model appearing in a platform catalog establishes availability on that platform under its conditions. It does not necessarily establish the model's original release date, its availability elsewhere, or feature parity across providers.

Before comparing prices, compare the same job. One product may include storage and tools while another charges only for model inference. Include retries, review time, tool calls, and any separate hosting costs in your own calculation.

5. Compare at the layer you intend to change

If you are replacing a model inside an existing application, keep the workflow and test inputs fixed. Compare task quality, latency, and cost.

If you are choosing an agent product, evaluate the complete task: setup, source access, intermediate decisions, approval controls, output quality, and recovery after failure.

If you are choosing a platform, inspect deployment options, access controls, export formats, observability, and how easily you can leave with your data and artifacts.

A strong model result cannot prove that an agent handles permissions correctly. A polished agent demonstration cannot prove that its underlying model is best for your classification task.

A brief you can reuse

Write one sentence for each of these before choosing a product:

• The job: what outcome does someone need?
• The inputs: which files and accounts are required?
• The model task: which part requires interpretation or generation?
• The workflow: which steps should be fixed?
• The agent's discretion: what may it decide for itself?
• The boundary: which actions require review?
• The evidence: how will you decide whether it worked?

Use the workflow planner to turn those answers into a project brief. Then evaluate the same task across tools.

Sources and further reading

These publisher resources illustrate the different layers; they are not independent performance comparisons.

• OpenAI's description of Dots illustrates an agent product with connected tools and ongoing work.
• Cursor's changelog shows changes to agent environments and artifacts.
• Runway's model catalog illustrates a platform exposing multiple model families.

Product documentation checked on 26 October 2026. The support-ticket workflow is an illustrative design, not a reported deployment.