Agents
An Agent is an AI member of the workspace, not a webhook or a human sign-in identity. It has a display name, avatar, Workers AI model, profile instructions, active or paused status, isolated conversation memory, and one stable Sandbox computer identity.
Requirements
The deployment needs healthy Workers AI, Durable Object, Workflow, R2, and Sandbox bindings. Agent Sandboxes use Cloudflare Containers and therefore require the Workers Paid plan.
The Worker may become reachable before the Sandbox image finishes provisioning. After a fresh deploy or upgrade, allow several minutes before treating the first failed Agent run as a persistent problem.
Create an Agent
You need Manage workspace permission.
- Open the workspace menu and select Workspace Settings.
- Open Agents and select Add agent.
- Enter a name, choose a Workers AI model, and write profile instructions.
- Optionally choose an avatar.
- Create the Agent.
The model picker stores a Workers AI model ID. The default runtime uses the deployment's AI binding and does not require an OpenAI or OpenRouter key. A deployment can set AGENT_MODEL as its default; each Agent profile may select another model supported by that binding.
Pause an Agent when it should remain configured but stop accepting new conversations or Tasks.
Use an Agent in Chat
There are two common paths:
- Mention the Agent in a Channel it has joined.
- Create a Direct Message containing the Agent and send a message.
In a one-to-one Direct Message, Discoflare creates or reuses a Thread for the Agent conversation. Replies in that Thread keep addressing the same Agent without another mention. Channel mentions and group Direct Messages stay in their source Channel.
The selected model can inspect image attachments only when it supports vision. A text-only model is told that the visual input was unavailable; it should not claim that it saw the image.
Follow a turn
Agent activity is visible beside the conversation:
| State | Meaning |
|---|---|
| Queued | The message was accepted durably and is waiting for execution. |
| Thinking | The Agent is producing the next response or action. |
| Tool | A tool or computer action is in progress. |
| Waiting for approval | A risky command is paused until an authorized member approves or rejects it. |
The Agent streams one editable response message as work progresses. An authorized member can stop an active turn. Approval is a control boundary for the pending risky command; it is not a general claim that every Agent result has been reviewed.
Memory and computer isolation
Conversation memory is separated by Channel or Thread. Each Task Run receives another isolated reasoning context, so concurrent work does not share a transcript accidentally.
Each Agent also has one stable Sandbox identity. A Sandbox container can sleep and its local disk can disappear, so Discoflare restores the Agent's latest /workspace checkpoint from R2 before use and saves a new checkpoint after mutating tools.
This makes the Agent's computer durable across container lifetimes, but it is not a permanent virtual machine.
Access boundary
Agents never receive a browser session or a human login. They are subject to workspace and resource authorization:
- an Agent cannot cross into a private Channel it has not joined;
- Mail actions require an explicit Mailbox grant;
- risky conversation commands pause for approval; and
- creating, configuring, discovering, and invoking Agents requires workspace-management authority in the current release.
An OAuth or social-login connection for a human does not give an Agent repository, shell, credential, or deployment access.
Assign work through Tasks
Use Tasks when the work needs a durable run, visible progress, cancellation, recovery, and retained outcomes rather than a conversational answer. Continue with Tasks.
Future repository connections, richer evidence, and production approval gates are described in the Agents roadmap.