Core concepts
Choose an agent, exchange messages, handle requests, and continue a conversation.
Use Console or a Service SDK to ask an agent for a project update. This walkthrough introduces Service through that conversation. For Harness and Harness UI, see Choose your path.
Choose an agent
An administrator gives you a role in a workspace, which contains your team's agents, other resources, and conversations. Your roles in the organization and workspace determine what you can use or change.
An agent combines a model, instructions, and tools. Example instructions: “Summarize project progress and cite the information you used.”
| Part | Role in the agent's work |
|---|---|
| Model | Interprets requests and produces responses or tool calls |
| Tool | Looks up a task, reads a file, or performs another configured action |
| Connection | Gives access to external tools through MCP or an app account |
| Environment | Provides the files and command execution used by tools |
| Memory | Holds project decisions the agent can read and update across conversations |
Start with a model and instructions. Before you ask the agent to look up real progress, add a tool or connection that reads your project data. Tools and connections explains the setup.
Start a conversation
Ask: “Summarize the project status and list the next two tasks.”
In Console, choose an agent and send your message. In an application, start a conversation through an SDK.
You see replies and tool activity as they arrive. While the agent works, you can add guidance or stop it. Compatible guidance joins the active run at its next model request; other messages stay queued.
Answer a question or approve a tool call
Answer questions and approve or deny tool calls in their conversation cards, then choose Submit responses after completing every pending item. Applications collect answers and submit a complete resume batch when ready.
Continue later
Open Sessions and choose Continue conversation, then ask “Which task should I do first?” Applications send the follow-up to the returned thread reference. Work continues when you disconnect, and Service stores the history for your return. Inspect execution opens the diagnostic view when you need to investigate a problem.
| Information | Purpose |
|---|---|
| Thread history | Continue this thread |
| Memory | Save and retrieve information across conversations |
| Environment files | Store documents and other working files for tools |
To share decisions across conversations, add memory to the thread or agent defaults. Environment storage and lifecycle determine how long working files remain available.
Understand execution when you need it
A session groups related threads, each with its own history and inbox. A run processes input using the agent's model and tools. An attempt is one execution of that run; recovery starts another attempt.
Console's Debug view shows these execution details. Guidance can join an active run; answering a waiting request creates a successor run.
Configuration changes create immutable revisions, listed under Versions. New messages use the default revision unless they select another. Active runs and approval continuations keep their original revision.
Next, try an agent in Console or connect your application. See Agents, threads and runs for revision selection and execution details.