Make every agent your own.

Start with intent, not a sequence of commands. Use chat to ask a question, describe an engineering outcome, or continue refining earlier work. The agent uses the active project as context so the conversation stays connected to the design.

Tools connect the conversation to action. An agent can choose from enabled capabilities to search the model, create or update objects, run analysis, and work with connected sources. Tool activity and results stay visible in the conversation so you can review what happened and decide what comes next.

Start with Auto when you want Davinci to decide. Automatic selection lets the agent choose an appropriate model for each request, so everyday work stays simple even as the task changes.

Choose explicitly when the tradeoffs matter. Select a model for faster responses, deeper reasoning, or more context, then adjust the options that model exposes. Your choice stays attached to the conversation until you change it.

Decide when to look beyond the model. Leave web search off when the agent should stay grounded in the conversation and active project model. Turn it on when the task benefits from current external information, such as standards, component data, published research, or other public sources.

Model selection changes the underlying intelligence; personas change its direction. Define a persona’s purpose and approach to give the agent persistent guidance about the role it should play, the priorities it should follow, and the way it should reason through a task.

Create specialized collaborators for recurring work. A systems reviewer can emphasize interfaces and risk, while a requirements analyst can focus on traceability and ambiguity. Teams can maintain multiple project personas and select the right one from the chat input.

Keep the model and persona independent. Switch models for speed, reasoning depth, or context size without losing the prompting that makes the persona useful, or apply the same model through several personas to support different workflows.

Skills capture instructions for repeatable work. Give each skill a name, description, and complete guidance so the agent can recognize when that expertise applies and follow the process consistently.

Categories help target the right kind of task. Organize skills around research, workflow guidance, or generation, then narrow generation skills to relevant model subjects and file types when needed.

Project skills keep shared practices close to the work. Teams can organize related guidance into folders, enable or disable individual skills, and refine instructions as their engineering process evolves.

Toolboxes let users create custom tools for the agent to use. Connect the agent to external data sources and services, or give it deterministic functions for analyses, transformations, calculations, and workflow automation.

Clear documentation tells the agent how each tool works. Define what a tool does, what information it expects, and when it should be called, then activate only the capabilities appropriate for the current project.

Test tools beside their implementation. Exercise individual actions with representative inputs, inspect their results, and refine the code before relying on the toolbox in an agent conversation.

Tools ground the agent in the active model. Instead of relying only on conversational context, the agent can retrieve project data directly, inspect relationships between model objects, and use current design information to guide its reasoning.

Explore, modify, and execute through a shared toolset. Agents can search the model, create or update objects, run scripts and analyses, and invoke other supported model operations. Each result becomes context for the next step, allowing the agent to work through a task without losing its connection to the design.

Tool activity remains visible. Calls and results appear in the conversation so you can follow the process, review changes, and decide how to continue.

Parallel conversations let independent work continue simultaneously. Start separate agents for analysis, requirements, verification, research, or other tasks instead of waiting for one conversation to finish before beginning the next.

Agent tabs keep every thread within reach. Each tab preserves its own conversation and settings, while visible status indicators distinguish agents that are working, queued, waiting for your input, complete, or unable to continue.

Shift your attention where it matters. Move between agents as results arrive, respond when one needs a decision, and return to earlier conversations through agent history without losing the context behind the work.

Every user message provides a clear restoration point. Use its checkpoint control when you want to return to that point in the conversation and reconsider the direction the agent took afterward.

Checkpoint restoration targets the agent’s model operations. Changes the agent applied after the selected message are undone without reverting unrelated editing that happened elsewhere in the project.

A restore remains non-destructive until you continue the agent. The later conversation stays available in a dimmed state, and Redo Checkpoint reapplies its model operations to return to the original result. Sending new guidance from the restored message commits the new direction and replaces the superseded path.

The Context Meter makes the active context window visible. Its live breakdown separates system instructions, tool definitions, user messages, agent responses, and tool results so you can understand what is being sent to the model.

Monitor both composition and capacity. The segmented meter shows how each context pool contributes to the total while tracking the tokens already used and the space still available for continued work.

Open the detailed view when the history matters. Follow how context changes across tool rounds and identify the sources consuming the window as a conversation becomes longer or more complex.

The compact meter stays beside the chat controls. It summarizes current credit consumption without requiring you to leave the conversation or interrupt an active workflow.

Expand the meter for a source-by-source breakdown. Review included usage, trial credits, and on-demand spending in credits or their configured monetary value, with progress bars showing consumption against available limits.

Know which balance is funding the work. The active usage source remains clear, helping individuals and managed teams understand where current agent spending is being applied.