Plan for build-time consumption
On the GitHub Copilot harness, credits accrue while you build and evaluate an agent, before it ever handles a production request. Some building activities consume credits and some don't. In this unit, you learn which build-time activities consume credits, so you can manage cost proactively instead of discovering it after launch.
What consumes credits before you publish
Three building activities consume Copilot Credits on the GitHub Copilot harness, and each one draws from capacity for the same reason: each one puts a language model to work.
| Activity | Consumes credits? | Notes |
|---|---|---|
| LLM-powered creation (natural-language authoring) | Yes | Runs a model to build or revise the agent, so it consumes credits; more iterative sessions consume more. |
| Preview conversations | Yes | Testing an agent in Preview runs the model, so it consumes like a real interaction. |
| Evaluation runs | Yes | Generating and running evaluations puts the model to work and consumes credits. |
| Manual configuration in Build | No | Setting fields and wiring configuration by hand doesn't invoke a model. |
The common thread is simple: an activity consumes credits when it runs a language model. LLM-powered creation is the natural-language authoring experience (in preview at time of authoring), where you describe what you want and a model builds or revises the agent, and Preview runs the same model and orchestration a published agent uses, so each consumes much like a live interaction. Evaluation runs consume too, as the agent works through the scenarios you define. More iterative authoring costs more, so thinking through what you want before you start authoring, rather than shaping the agent through trial and error, is itself a cost lever.
Manual configuration is the exception. Setting instructions, wiring up knowledge sources, adding tools, and adjusting fields in the Build tab doesn't run a model, so you can structure an agent for hours without drawing a single credit. Consumption starts only when a model runs.
Use build-time consumption to estimate production cost
Because Preview and evaluation run an agent the same way production does, the credits they consume give you an early read on what those same scenarios cost once the agent is live. Testing against the requests an agent handles in practice makes build-time consumption a directional signal of production cost. Treat it as a signal, not a billing guarantee. Pair it with the Copilot Credits estimator for the GitHub Copilot harness to plan capacity before you publish. After you publish, the Monitor tab shows real consumption to compare against your estimate — the same consumption administrators monitor across the tenant.
Reflect: Think about an agent you're about to build. Which representative scenarios would you run in Preview or evaluation to get an early read on what it's likely to consume in production?
Whether credits accrue while you build or while your agent runs, they draw from the same tenant capacity. The next unit turns to how administrators fund and govern that capacity.