Subagents & Sandbox

Two pieces of machinery set the Dynamiq Agent apart from a plain chatbot: a sandbox — a real cloud computer provisioned for each conversation — and subagents, parallel helper agents the main agent can spawn for large tasks. Both are part of Dynamiq Agent mode.

The conversation sandbox

When you start a Dynamiq Agent conversation, the platform provisions an isolated cloud sandbox for it. The monitor icon with a green dot next to the model selector — Computer is running on hover — confirms it is live.

The sandbox is where the agent's hands-on work happens:

• Terminal — the agent runs shell commands; each one appears as a "Computer terminal" step you can open in the tool details panel.
• Files — the agent's working directory is /home/user; files you attach land in /home/user/input, and deliverables it writes there surface as downloadable cards in the reply.
• Code — Python scripts, data processing, document generation — anything runnable from a shell.
• Web serving — sites the agent builds can be previewed live in the web preview panel.
• Skills — your enabled Skills are mounted into the sandbox for the agent to read.

The sandbox belongs to the conversation. Come back to the conversation later and the platform reattaches to it — or transparently provisions a fresh one if the old sandbox has expired (a fresh sandbox starts with an empty filesystem, but your conversation files remain downloadable from the chat).

Subagents

The branching-arrows toggle in the input bar enables subagents — hover shows Enable subagents for large scale research. When it is on (filled), each message you send lets the main agent delegate work to subordinate agents.

A subagent is a full agent in its own right: it gets a fresh, isolated sandbox plus web search, scraping, and browser tools, and runs independently of the main conversation. Because each invocation is isolated, the main agent can run several subagents in parallel — you'll see "Sub-agent" steps fan out in the reply.

The main agent is instructed to use them deliberately:

• Good fits — deep research across many sources, fan-out (the same operation over many independent items), reading large document sets without bloating the conversation's context, and independent verification of finished work.
• Not used for — small tasks (a few tool calls), reformatting things already in context, or anything that needs your conversation history: a subagent sees only the task description it is handed.
• Limits — the agent aims for at most 5 subagents per task and is capped at 20 subagent calls per message; the main agent keeps final synthesis and assembly to itself.

Subagent results flow back to the main agent: files a subagent produces are saved into the conversation's own sandbox automatically, and the main agent reuses them in the final deliverable instead of regenerating the work.

For engineers: the toggle maps to the optional subagents_enabled boolean on Send Conversation Message. The equivalent building block for your own workflows is the Sub-Agent Tool — see Subagents and Delegation.

Next steps