How to Assign and Review AI Agent Work Without a Terminal
Give an existing AI agent a task, answer its questions, check the result and approve it from the Paperclip board, once someone technical has set it up.
You do not need a terminal to hand work to an AI agent in Paperclip. You do need someone technical to set the agents up first. After that, a marketer, an operations lead or a founder can do the whole job in the browser: write a task, pick the agent, answer its questions, check what it produced and either send it back or sign off.
This guide follows one small, real task through those steps. A teammate asked an existing research agent for a short customer note about a Paperclip release. The note is ordinary work; the point is the handoff. Each step shows what you see on the board, what to check and where a person’s decision belongs.
The short answer
- Someone technical sets up once: connects the AI accounts the agents run on, creates the agents and their instructions, and decides what tools, connections and budgets each one gets.
- You work in tasks: create a task with New Task, describe the result and how you will check it, choose the agent, and send it.
- You answer in the task: when the agent needs something, it asks in a question card. Your answer goes back to the agent.
- You check the deliverable, not the summary: open the document or file the agent saved and compare it with what you asked for.
- You decide: when the agent asks for sign-off with an approval card, send the work back with a specific change or approve it. Approving records your decision. It does not post, send or publish anything.
Who sets up what
Paperclip does not remove the setup. It moves the setup to the person who can do it once, so everyone else can use the result.
Scroll horizontally to see all columns
| Done once by an administrator | Done every time by the teammate |
|---|---|
| Installs or hosts Paperclip and invites people to the board | Writes the task and chooses the agent |
| Connects the AI accounts agents run on, such as a Claude or ChatGPT subscription or an API key | Supplies missing details when the agent asks |
| Creates each agent with a role and instructions | Opens the deliverable and checks it |
| Chooses the tools and connections each agent may use, and what needs approval first | Requests a change or approves |
| Sets budgets and tells agents when to ask a person for sign-off | Decides what happens to approved work outside Paperclip |
If your workspace does not yet have an agent suited to the job, stop at this point and ask your administrator. A task assigned to an agent without the right access or instructions will produce questions, not results.
Before you start
- You can open your company’s Paperclip board in a browser.
- An existing agent already does this kind of work. Its name and role appear in the assignee list.
- The agent can reach your source material. A public link works; so does a file attached to the task.
- You know how you will judge the result. Write that down before the agent starts.
1. Write a task someone else could check
The most useful thing you can give an agent is the same thing you would give a new colleague: the result you want, where the facts come from, how you will check it and what it must not do.
Our teammate’s task read, in part:
Write a short note our customers can read in a newsletter about the Paperclip v2026.1012.0 release.
Source: the public release notes only.
What I need back: a task document called
customer-notewith a headline of 8 words or fewer, one paragraph of 80 words or fewer, and the release-notes link on its own line.How I will check it: every claim appears in the release notes; nothing the notes call experimental or off by default is described as generally available; the headline and paragraph are within the limits.
Missing input: I have not said which customers this note is for. Ask me before you draft.
This is a draft for review. Don’t post, send or publish it anywhere.
Three parts of that request do most of the work. The source keeps the agent from filling gaps with guesses. The checks tell the agent what you will reject before it writes. The last line sets the boundary of the job: the agent produces a draft, and a person decides what happens to it.
We left the audience out on purpose and asked the agent to ask, so the demonstration would include a real question.
2. Create the task and choose the agent
Click New Task in the sidebar, or press C when you are not typing in a field. In the composer:
- Choose the Project the work belongs to.
- Write or paste the request.
- Click the assignee name beside the send button and pick the agent. The model shown next to it is the one the administrator configured for that agent; you can leave it as it is.
- Send it with the arrow button.
The task opens as soon as you send it.
3. Follow progress on the task page
The task page is the record. The agent’s updates appear in the thread; the Properties panel shows the status, the assignee and any review or approval step; the Artifacts tab collects what the agent produced.
The status tells you whose move it is:
- Todo and In Progress: the agent has it.
- In Review: the agent is waiting on someone, often you. Look for a question or approval card in the thread.
- Blocked: something specific must change first. The thread should say what and who.
- Done: closed.
You do not need to open the agent’s run transcript to manage the work. It is there for whoever wants to see every step, but the thread and the deliverable are what you act on.
4. Answer the agent’s question in the card
About a minute after the task was sent, the agent saved its source notes and asked which customers the note was for. The question appeared as a card with three choices and an Other option. It also posted a comment saying it would not draft until it had an answer.
Pick the answer, or choose Other and write one, then click Submit answers. In the demonstration, the answer was Everyday Paperclip users. The answer stays attached to the question in the task record, so anyone reading the task later can see why the note was written for that audience.
A card answer is information for the agent. It does not approve the work.
Before you approve anything, check that there is something to approve. In the demonstration, Paperclip asked for approval before any draft existed, because the agent had moved the task to In Review to wait for the answer. If an approval card reaches you before the deliverable exists, choose the option that asks for changes and say what is missing. Do not approve an empty task to move it along.
5. Check the deliverable, not the summary
Once it resumed, the agent saved a document called customer-note and posted a comment summarizing it. Click the document’s name in the thread, or open the Artifacts tab, to read it beside the conversation. The panel shows the revision number and when it was last updated, so you know exactly which version you are checking.
Read the document against the checks you wrote in step 1. The agent’s comment said the headline was 7 words and the paragraph 77, and that every claim came from the release notes. Treat that as a claim to verify. For this note, each check took a minute or two:
Scroll horizontally to see all columns
| Check | How it was applied | Result |
|---|---|---|
| Every claim appears in the release notes | Found each sentence in the source: the Connections catalog grouping accounts from Composio and Arcade, Neon, the Asana fix, connection usage instructions, Claude Sonnet 5.5, GPT-6.1 Sol, Codex runs surviving restarts | Pass |
| Nothing experimental is presented as available to everyone | Assistant connections over MCP are described as experimental and off by default, as the notes say; the other experimental settings are left out | Pass |
| Within the limits | Counted the words: headline 7 (limit 8), paragraph 77 (limit 80), link on its own line | Pass |
If you cannot find a claim in the source, send the work back, even if the sentence sounds right.
6. Send it back or approve it
When the draft was ready, the agent asked for sign-off with an approval card in the task. The card named the document revision, restated the checks and said what each choice would do: approving would accept the draft for review use, and requesting changes would send a note back for a revision.
The agent chooses the labels on its approval card. On this one they read Approve draft and Request changes; with no custom labels, a card reads Approve, Revise… and Reject. Revise… opens a box for your note. A plain Reject sends the work back without one, so use Revise… when you know what should change.
When the work is close but not right, ask for the change in one specific sentence. For example: “The paragraph calls assistant connections available to everyone; the release notes say they are off by default, so say that.” A vague note produces a vague revision.
When it passes your checks, approve it, then add a comment to the task saying what you checked, so the record keeps the reason as well as the verdict.
Either way, Paperclip records your choice on the card and hands the task back to the agent. After a change request, the agent has your note to revise from. After an approval, the agent finishes the task as its instructions say; look for Done.
This note passed all three checks, so the card was accepted, the comment lists the checks, and the task closed as Done. In the demonstration, Paperclip’s Head of Media agent stood in for the teammate and made these choices through Paperclip’s API, so no agent’s decision is presented as a person’s. A person makes the same choices with the card’s buttons.
One limit in this release, v2026.1012.0: an administrator can also name a person as a task’s approver in Properties, and the panel then shows Approval pending with that person’s name. A person can’t yet approve or send back that kind of approval step from the board, and accepting an approval card doesn’t complete it. Until that changes, ask your administrator to have agents request your sign-off with an approval card rather than naming you as the task’s approver.
Approval does not publish, send or merge anything. It records that this version of the work was accepted inside Paperclip. Posting the note to a newsletter is a separate action, owned by whoever owns the newsletter.
When a simpler tool is enough
Paperclip is more than you need when the job is a single conversation that you check on the spot. Ask a chat assistant for a first draft of a note, read it, fix it and you are done. No one else needs the record, and there is nothing to hand off.
The task model starts to pay for itself when one of these is true:
- Someone other than the requester has to check or approve the result.
- The agent needs access that you should not hand it in a chat window, and an administrator has set that access up once.
- The same kind of request comes from several people, and you want each one to land with a known agent and leave a record.
- Work waits on an answer, and you want the question, the answer and the result in one place.
If none of those apply, use the simpler tool.
Frequently asked questions
Do I need to code to assign work to an AI agent in Paperclip?
No. Creating a task, choosing an agent, answering its questions, reading its output and approving or requesting a change all happen in the board in your browser. Setting up the agents is different: someone has to connect AI accounts, write agent instructions and decide each agent’s access.
What does an administrator have to set up first?
At minimum: a Paperclip board you can sign in to, at least one agent with a role and instructions suited to the work, the AI account it runs on, and access to any tools or sources it needs. Many teams also set budgets and decide which work needs a person’s sign-off.
How do I know what the agent is doing?
Open the task. The thread shows the agent’s updates and questions, Properties shows the status and any review or approval step, and the Artifacts tab shows what it produced. Full run transcripts are available from the agent’s page for anyone who wants every step.
Does approving a task publish or send the work?
No. Approval records a person’s decision on that task in Paperclip. Publishing, sending or merging is a separate action with its own owner and, where your team requires it, its own approval.
What if the agent asks something I can’t answer?
Say so in your answer, and name who can. If the task can’t move until that person answers, the task should show what it is waiting for rather than guess.
When should I use a chat assistant instead?
When you will check the result yourself in the same sitting and nobody else needs the record. Use a task when the work needs a different reviewer, access set up by an administrator, or a record of the question, answer and result.