Sparse Agent can keep working toward a goal you set. You describe the result you want, and the Agent advances the work through subsequent tasks. Return to the workspace whenever you need to add instructions, review progress or help with a sign-in.
For a first goal, choose something with a clear scope and an outcome you can check: preparing project updates, following a topic or managing an account. This guide uses Weibo operations to show how to complete setup, define a goal and start working together. Adjust the example timeframe and numbers to your own plan.
Sign in and complete setup
Open the Sparse Agent workspace and sign in with Auto-Coder Chat. Existing members can sign in directly. First-time members should use an invitation code when prompted.
The setup guide takes you through Model → Cloud key → Create Agent. You can also skip it and complete these steps later in Settings:
- Model. Register with your chosen model provider and create an API key. Select the matching provider in Settings → Model, enter the key and save.
- Cloud key. Register your own DigitalOcean account, enable billing and complete any requested deposit or payment verification. Generate an API token with read and write permissions and save it in Settings → Cloud key. DigitalOcean is the default. Alibaba Cloud ECS and private Kubernetes are under More configuration on the same page, and private Kubernetes needs a container image reference.
- Cloud environment. Create the computer on the environment page or as part of New Agent, then wait for the Ready status. New environments have no time limit by default; optional automatic deletion can be set for 1–168 hours at creation.
- Cloud memory (recommended). Authorize the current Chat account in Settings → Cloud memory so later tasks can use saved preferences and experience.
The model and cloud machine use your own provider accounts and their billing terms. Powered-off cloud machines can still incur charges; delete the environment and confirm completion when finished. The four-settings guide explains registration, payment, token permissions, each setting's role and why Omarchy and DigitalOcean are used.
Enter keys only in the relevant settings fields. After basic setup, you can also connect cloud memory to help the Agent retain your working preferences, as described below.
Goal: name the account, work and timeframe
Give the Agent a recognizable name and fill in the required Goal. Describe the work you want it to carry out over time, including whom it serves, what it should do and the timeframe.
“Run my Weibo account” leaves several details open. You could write:
Goal
For the next two months, operate our product's Weibo account to help more potential customers learn about the product.
Publish at least two posts a day based on real source material, choose topics and review feedback. Send me the day's operations report after 21:00.
The Agent now knows whom it serves, its daily duties and how you want to hear about progress. If the scope is still unclear, ask it to review the current situation and make recommendations before deciding on the ongoing work.
KPIs: explain how to measure progress
Expand KPI and Constraints (optional) in the creation form. Use KPIs to specify a target, the data source, the counting method and how often to update it.
The 100,000 target below is an adjustable example:
KPI
Reach 100,000 verifiable monthly views, using the post read counts in Weibo's dashboard. Record each post and add the readings together.
Count each post's reading once. Record likes, comments and repost counts separately.
Update daily. When fresh data is unavailable, mark the previous reading as carried over and retain its measurement time.
Two daily posts describe the work to do; the view target measures its result. Review both to see whether work continues and whether it produces the expected outcome. Views count readings: one person reading different posts may count more than once.
Constraints: say when to ask you
Use Constraints to describe the limits of the work and where you need to participate. For example: “Send drafts to me for approval before publishing. Ask for help when a sign-in is required. Do not take paid actions without approval.”
Keep these instructions consistent with the Goal. If the Agent may publish independently, specify the permitted accounts and content. If you want to review drafts first, say that publication requires approval. Add new requirements as they arise.
Create the Agent and start a task
Check the name, Goal and optional requirements, choose an environment as prompted, then select Create Agent. Follow the preparation progress on the page; it remains available after a refresh.
New environments have no time limit by default. They stay until you delete them manually and continue to incur cloud charges. To schedule cleanup, fill in Auto-delete (hours, optional) at creation with 1–168 hours (up to seven days); leave it blank or choose No time limit to keep the environment without an expiry. An environment with a configured expiry is deleted when it expires, so save the work you need beforehand. Existing environments keep their previously configured expiry.
Once preparation finishes, choose the Agent on the left, check its status and send a specific task. For example: “Review last week's posts and suggest topics for today. Wait for my approval before publishing.” Replies and previews remain in the conversation.
Review results and adjust the next steps
Open the Agent's details to view Goal and progress and Service status. Assess completed work, delivered material and KPIs together. “Goal created” alone does not mean that a task is running or complete.
Send a message to add context or assign the current task. To change ongoing requirements, use Edit Goal / KPI in the details, update the Goal, KPIs or Constraints and save. Sending a new message does not automatically replace the original Goal.
Use Pause when you need to stop temporarily and Resume when you are ready to continue, then check the updated status. Available buttons depend on the task's current state. If the Agent needs you to sign in or provide other help, open Needs your help and follow its instructions before continuing.
Use email and Telegram through the browser
Sparse Agent can operate websites through its browser. After you sign in to your account in the browser used by that Agent, you can ask it to work through that account: send email, set up message notifications or carry out other actions on a website.
Ask the Agent to open the website, then use Browser → Preview to see the current page. If a QR code, verification code or sign-in confirmation is required, follow the Needs your help prompt and tell the Agent to continue. Sign in within the browser the Agent is using.
Send email from a signed-in mailbox
To have the Agent email a work report to another address, ask it to open your webmail and help it complete sign-in. Then specify the recipient, subject and content. It can send the email from that signed-in mailbox.
Use the mailbox I just signed in to through your browser to send today's work report to the recipient address I specify. Use “Project daily report” as the subject and tell me whether sending succeeded.
Start with a simple email, check the sending result, then arrange subsequent reports. If you want to review a draft first, add: “Show me the content for approval before sending.”
Set up Telegram messages on your phone
You can also ask the Agent to open Telegram on the web, sign in to your Telegram account in its browser and have it help set up Telegram notifications. Complete any scan or confirmation steps when prompted.
Help me set up Telegram notifications so task updates reach Telegram on my phone. Once setup is complete, send a test message. After I confirm receipt, use it for subsequent daily reports.
After checking that the message reaches your phone, agree on what to send and when: daily progress, completed task results or items that need your attention. You can then receive work updates on your phone without keeping the workspace open.
Use earlier experience in later tasks
In Settings → Cloud memory, select Connect Chat account, authorize the connection and check its status. Start with a preference you can verify, such as “Put results first in daily reports, then list unfinished work and next steps.” Later, ask the Agent to retrieve and apply it.
The connection applies to subsequent tasks for existing and new Agents. To bring existing memory into the cloud, choose Migrate existing memory on the same page and follow its progress. The original local records are retained.
See the cloud memory guide for connection, query and update examples. For viewing and backing up memory on the web, read the Sparse Agent Memory guide. Cloud memory retains reusable information; save important deliverable files separately too.
Send follow-ups from another Agent
CLI and MCP connections let you send instructions and read progress from Codex, Claude Code or WorkBuddy using your Sparse Agent's existing goal, memory and conversation.
CLI / MCP access is enabled on the official website. Sign in and open Settings → CLI / MCP. See the third-party Agent access guide for installation, authorization, requests and reconnecting. Connected clients share the same chat history with the website.
Check these points on your first run
- Model and cloud services are configured, and the Agent's environment is available.
- The Goal names the account, work and timeframe, and the KPIs explain how to measure progress.
- The Agent has received its first task, and you can review its reply and actual results.
- Approval requirements, reporting expectations and environment expiry are clear.
You can expand the scope as you gain experience. To see a real goal compared with its results, continue to the Weibo operations case.