Deep agents

Every agent runs on a harness, chosen under Deep Agent Mode on the agent form.

Harness	Best for
Classic — standard conversational agent	Answering questions, looking things up and calling a few tools in a turn. Faster and cheaper per turn.
Deep — planning, skills, memory & sub-agents	Multi-step work: research, drafting documents, working with files, tasks that need a plan and several tools in sequence.

Start with Classic. Switch to Deep when the agent needs to plan, keep notes between conversations, or follow skills.

What Deep adds

• Planning. The agent breaks a request into a to-do list and works through it step by step.
• Skills. Playbooks the agent opens only when a task matches them. See Skills.
• Long-term memory. A notes file the agent reads at the start of every conversation and can update as it learns.
• Sub-agents. The agent can hand parts of a task to other agents. Attach them in the visual canvas.
• Files and commands. A workspace where the agent can write files and, when a sandbox is available, run shell commands.

Memory

With Deep selected, the form shows two memory fields.

Field	What it does
Initial memory seed	Text loaded into the agent’s memory the first time it is used, for example “The company sells solar panels. Always respond in Spanish.” The agent can update it over time.
Current memory	What the agent has stored so far, with the time of the last update. Reset memory clears it; this cannot be undone.

Memory is capped at 16 KB.

Caution

There is one memory per agent, shared by every conversation and every user of that agent in the workspace. Don’t let an agent store personal details it should not repeat to someone else; say so in its system prompt if needed.

Where commands run

When a deep agent runs a shell command, the command runs in the first of these that is available:

1. A paired VirtuAI CLI for the workspace: the user’s own machine, or a container running the CLI.
2. A cloud sandbox, when the workspace has an E2B_API_KEY in Workspace settings.
3. A virtual file system otherwise. The agent can still read and write files, but it cannot run commands.

In web chat and the CLI, each command waits for the user’s approval first. See Command approvals.

Where Deep applies

Deep mode applies in web chat, the CLI, Google Chat, Telegram, Slack and A2A. WhatsApp and voice run the agent as Classic.

Before you switch

• Deep agents make more model calls per request, so they cost more and take longer. Watch cost in Analytics after switching.
• Give the agent a Max Output Tokens large enough for the documents you expect it to write.
• Re-run your evaluations after the switch. A deep agent can pass or fail cases differently from the same prompt on Classic.