Chat, approvals and previews
Talking to an agent
Type in the box at the bottom and press Enter. The chat looks the same whichever agent program is behind the agent: your messages, the agent's replies, each tool it uses, and any request for permission.
An agent can hold several conversations. Many settings take effect only when an agent is restarted, and a restart does not end a conversation: write in the same conversation afterwards and the agent carries on, remembering what was said. If an agent program cannot pick a conversation up again, the history stays readable and your next message starts a new one.
The agent program's own commands
Agent programs understand short commands typed at the start of a message, such as /compact to shorten a long conversation or /model to switch models. Type / in the message box and a list of the commands that agent's program understands appears above it. Keep typing to narrow the list.
| Key | What it does |
|---|---|
| Up and down arrows | Move through the list |
| Tab or Enter | Put the chosen command in the box, ready for anything you want to add after it |
| Enter, once the box holds the command and anything after it | Send it |
| Escape | Close the list |
A command is sent like any other message, and the agent program does what it says. Its answer appears in the chat.
Each program has its own commands, so the list differs from agent to agent. It also includes the skills and plugins that agent has. The list appears once the agent has opened its first conversation. Codex writes its skills with a $ in front of the name, such as $imagegen, and the list puts them in the box that way.
Some commands only make sense in the program's own terminal window, for example ones that change its colours. They are in the list because the program lists them, and they may do nothing useful in the chat.
What an agent is doing
| Shown as | Meaning |
|---|---|
| Ready | Running and waiting for a message |
| Working | In the middle of a task |
| Waiting for you | It asked permission for something and is waiting for your answer |
| Stopped | Not running |
| Stopped with a problem | Its agent program stopped by itself. The reason is shown |
Approvals
When an agent wants to do something that needs permission, it stops and asks. The request appears in the chat, on the Home page, and in the Inbox, which has the answer buttons, so you can answer without opening each agent.
Each agent has one of three settings, in its Settings under Safety:
| Setting | What happens |
|---|---|
| Ask every time | The agent waits for your answer before any action that needs permission |
| Allow reading, ask for the rest | Looking at files and searching go ahead. Changing files or running commands waits for you |
| Allow everything | The agent never waits |
A change takes effect the next time the agent starts.
If DAC itself restarts
Agents do not depend on DAC staying up. A task in progress carries on while DAC is stopped or being upgraded, and the dashboard shows everything that happened once DAC is back. An agent that needs permission waits until then.
This covers DAC restarting, not your computer rebooting. After a reboot the history is still there, but a task that was in progress is lost.
Previews
When an agent starts a web server, for example to show an app it is building, the server appears under Previews beside the chat as a link such as "Open port 21000". Pressing it opens what the agent is serving.
Each agent has its own block of twenty network ports and is told the first one, so servers started by different agents do not collide. DAC cannot force a program to stay inside its block, so a program that insists on a fixed port can still clash with another.
Instructions and settings
Every agent program has a file of standing instructions, which the agent reads at the start of every conversation. Each agent has its own copy, which you can change in its Settings under Harness. This is how two agents on the same program are told different things, for example one to act as a reviewer and one as a writer.
A new agent's instructions already ask it to write to you plainly: the main point first, every step of its reasoning shown, and technical terms explained. You can change or remove that text like any other.
The agent program's own options
Each agent program has options of its own, such as which model to use. They appear in the agent's Settings under Harness once the agent has opened its first conversation. A choice belongs to one agent, so two agents on the same program can use two different models.
The Terminal tab
Each agent has a Terminal tab: a command line running inside that agent's own environment, with its home folder and project folder. Ordinary use never needs it. It is there for things that only work in a terminal, such as an agent program whose sign-in has no button in the dashboard.
Stopping and deleting
Stopping or deleting an agent stops everything it started, including programs it left running in the background. Another agent's programs are never touched.
Deleting an agent removes its files
Delete removes the agent's whole folder: its settings, connections, skills, conversations and the files in its project folder. Push any work you want to keep to its repository first. The dashboard reminds you before you confirm.