Harnesses and models
Your agent's job stays recognizable from one run to the next, but something still has to interpret the request, use the tools, and connect to a model provider. Agenta separates those runtime choices into a model, a harness, and a credential.
One agent has three runtime choices
Return to the marketing launch agent that prepares the Monday campaign report. Its instructions describe the audience. Its reporting skill describes the process. Its tools read analytics and post the final summary.
To run that agent, you make three additional choices:
- The model works with the language and makes the next decision.
- The harness keeps the task moving between the model, tools, files, and conversation.
- The credential connects the harness to the place serving the model.
These roles work together, but they solve different problems.
The model provides the intelligence
The model reads the context it receives and produces the next response. For the campaign report, it may decide which results matter, explain why a metric changed, choose an available tool, or draft the final summary.
Models differ in quality, speed, cost, context size, and the kinds of inputs and tool use they support. A model that is a good fit for a short support reply may not be the model you choose for a long research task.
Changing the model changes how the agent reasons and writes. It does not rewrite the agent's instructions, skill, or files.
The harness keeps the task moving
The harness is the software that runs the agent around the model. It prepares the context, sends work to the model, carries out supported tool requests, returns the results, and continues until the task reaches an answer or another stopping point.
For the Monday report, the harness might:
- Give the model the request and the agent's working context.
- Run the analytics tool the model chooses.
- Return the campaign data to the model.
- Pause if posting the report needs approval, or run the posting tool when it is allowed.
- Keep the resulting messages and tool activity together in the session.
The harness also manages practical details such as loading instructions and keeping a long conversation within the model's limits.
Agenta offers two harnesses, Pi and Claude Code. They support different model providers and runtime capabilities, so the harness you choose affects which models and features are available to the agent.
The credential connects to the model provider
The credential answers a separate question: how does this agent authenticate with the service running the model?
That may be:
- An API key for the model provider.
- A project connection to an OpenAI-compatible endpoint or company gateway.
- A cloud-hosted model connection supported by the selected harness.
- On self-hosted Agenta, a supported Claude or ChatGPT subscription login available to the harness.
The model name alone is not enough. The credential tells Agenta where that model is served and how the run is authorized.
You can change one choice without redefining the job
Suppose the launch agent produces a good report, but you want to compare a faster model. You can change the model and test the same request again. The audience, report process, campaign files, and delivery rules remain part of the agent.
You can also change the credential when your company moves model traffic behind its own gateway. That connection change does not require you to rewrite the agent's marketing instructions.
Changing the harness is a larger compatibility choice because harnesses support different models and capabilities. Review the supported model and hosting options, then test the agent's tools and behavior before committing the change.
| Choice | It decides | Change it when |
|---|---|---|
| Model | How the agent interprets and generates the next step | You want a different quality, speed, cost, or capability tradeoff |
| Harness | How the agent runs around the model | You need a different runtime or supported provider set |
| Credential | Where model access comes from | You change provider keys, gateways, hosting, or supported subscription access |