Agent Workbench

#Two workbenches separate AI collaboration from manual operations

Affiliate provides two workbenches with different purposes:

• Agent Workbench: handle the points where AI work needs a person and inspect the full Agent work record. It contains Pending review, Pending, and All Agent work.
• Manual Workbench: centrally process Shop back-office tasks that belong to people from the start. It contains Sample Applications and Creator messages.

This separation keeps three cases distinct: an action caught by an Approval Policy enters Pending review; an issue actively escalated by the Agent enters Pending; work for protected Creators or Human-only BDs bypasses AI and is handled by staff in the Manual Workbench.

With an individual BD login, Agent Workbench offers My Creators and Public Creators. Manual Workbench can show All, Public, or a specific BD's Samples and messages within assigned regions. Authorized daily operations are available for own/public Creators; colleagues' records are read-only. Reading, daily-operation, and Proposal-review permissions differ. See Team & Channels.

APP CAPTUREThe English light-theme Agent Workbench with Pending review, Pending, and All Agent work.

#First, distinguish review from handling

Both tabs require a person, but they represent different kinds of human work:

In short: review an AI-submitted action; handle an issue the Agent actively escalates. Review is not an Agent request for help, and handling is not simply approving an already-prepared action.

#Agent Workbench

Open Affiliate → Agent Workbench. Confirm BD and date scope, then choose a tab: review AI actions, handle Agent requests for help, or audit all AI actions. These are not three copies of the same queue.

#1. Pending review: review an Action Proposal

This tab contains operations the Agent has already decided and submitted for execution, but the policy system has held as Action Proposals. The item might be a message, a Sample application decision, a tag change, or a “no action needed” decision.

Account-level policy determines what enters this queue. The Agent does not decide whether a policy matches and does not need to know that review will be required. After the Agent submits the action, the system independently checks the Safety Policies under Team & Channels. A match holds the Action Proposal under Pending review; without a match, the operation can execute automatically.

Review in this order:

1. Subject: confirm Creator, Product, Shop, and Campaign.
2. Evidence: compare the Agent’s reason with relationship history, Product Knowledge, and application data.
3. Action scope: verify that the message, benefit, Sample decision, or tag is authorized.
4. Language and risk: check for unsupported promises, sensitive wording, or incorrect facts.

Approve when the action is ready, reject it when it should not run, or request a revision. This queue asks a person to review a prepared Action Proposal—not to handle the entire Creator case from the beginning and not to respond to an Agent escalation.

Generic message/tag Proposals offer Request revision and Ignore. Supply specific feedback before submitting a revision request, then inspect the new version. Ignore asks for confirmation and drops that action without performing the platform operation. This differs from the business decision to ignore a Sample application below. For a multi-action Proposal, inspect every action rather than just the first message or Product.

#Review a Sample action

When policy holds a Sample action, open the proposal to compare the Agent recommendation with the application, Creator performance, and relationship context.

APP CAPTUREThe English light-theme Sample action review with the Agent recommendation and evidence beside the human decision controls.

Samples have real cost. Check Product fit, previous publishing and order performance, the current Campaign offer, Sample Tier, Manual Tags, and protection rules. Approval continues the prepared Sample operation; rejection records the human decision and ends that operation.

When approving, rejecting, or ignoring, you can add an optional Review note for inventory checks, Creator fit, exception approval, or other reasoning. The note remains on the Sample record so the next staff member can understand the decision.

Two similar-looking actions have different effects:

• Reject: submit a real merchant rejection to TikTok.
• Ignore: remove the application from active review without rejecting it on TikTok, allowing it to expire naturally; it can be reopened while the platform still permits review.

When the Agent proposes ignoring a Sample, All Agent work labels the proposal as Ignore rather than describing it as a platform rejection.

#2. Pending: handle an Escalation

Pending is comparable to an AI Customer Service escalation. The Agent recognizes that the issue exceeds its current capability, tools, or authority, so it actively creates an Escalation and asks a person to take over.

For example, if a Sample is lost in transit, someone may need to decide whether to reship, create new fulfillment, or coordinate directly with the Creator. The Agent can organize the context and hand off the case, but it cannot perform an unauthorized reshipment on its own.

APP CAPTUREThe English light-theme Affiliate Pending list with open items that require human intervention and their related context.

To handle one:

1. open the row and read the Agent question, reason, and context, confirming the Creator and responsible BD;
2. perform the necessary manual operation or make the required decision;
3. use the single answer box to state your decision and what the Agent should do next, then submit; submission resolves the Escalation rather than saving an unfinished draft;
4. when the issue repeats, add reusable guidance to Business Instructions, Product Knowledge, or policy.

Unlike Pending review, there is no prepared action waiting to run after a simple approval. A person must understand the issue, make the decision or perform the actual operation, then record the outcome and resolve the Escalation.

For example: “I confirmed with the warehouse and arranged the replacement, tracking number … . Tell the Creator the new estimated arrival and follow up on delivery.” Do not write “replacement arranged” if you have not done it. Expressing an intention in this answer does not itself create a warehouse order.

APP CAPTUREEnglish light-theme Escalation detail with the Agent question, reason, context, and human answer.

The list's notification status describes delivery to the configured team channel, not resolution of the issue. You can still open and handle it here when notification delivery fails. See Team operations for BD notification routes.

#3. All Agent work: inspect the full work record

This tab aggregates Agent work across states and supports filtering by owner, status, type, or search. Use it to audit approved, executed, rejected, ignored, or expired actions and to reconstruct what happened for a Creator or Campaign.

It is not a duplicate of Pending review. Pending review contains only Pending Proposals currently waiting for a human decision; All Agent work provides the broader history. It is not another manual task queue.

#Check whether an action actually succeeded

Narrow by BD, date, type, and search, then select a status. Dates use Proposal creation time; the time selection is retained when switching between the three Agent tabs.

Types distinguish messages, Sample decisions, tag changes, and no-action decisions. Open a record for the Agent's reasoning, exact action, reviewer, and execution result; inspect feedback and revision history when present. Identify an ignored Sample from the actual action, not from a generic rejected status.

#Review does not expand permissions

BDs, supervisors, and account owners have different viewing and review scopes. Visibility does not grant approval rights. Reassignment can leave an existing Proposal under its original review scope. See the permission table.

If the state changed, refresh and inspect it instead of repeatedly approving. Another employee may have acted, or the business state may no longer allow the original action. After an execution failure, inspect the error and platform state before trying anything that could duplicate a Sample or message.

#Manual Workbench

Sample decisions, Creator replies, status filters, cross-Shop Product lookup, and attachments are documented separately in Manual Workbench. Pending here means an AI Escalation, not an original message waiting for a staff reply.

#Trace who acted, reviewed, and sent a message

Samples and the timeline identify the actual operator. All Agent work shows who decided a Proposal, and messages identify the sender. The responsible BD is the relationship owner, not necessarily the person who performed an action. Analytics Details also supports reviewer filtering.

After reassignment, existing Proposals retain their original reviewer scope rather than automatically moving to the new BD. Ask a supervisor to coordinate read-only colleague records or inaccessible old Proposals; repeated reassignment does not replace permission checks.

When no reply is required, open the conversation and select No reply needed · Mark handled to finish the human task without sending anything to the Creator. If the same message still has a Pending Proposal, decide that proposal first so the system does not keep a prepared action while the message is marked complete.

#Open Creator Relationship when more context is needed

A single work item shows one moment. For a costly Sample, long-running collaboration, reshipment, or unusual request, open Creator Relationship to inspect cross-Shop interactions, Campaigns, Samples, Collaborations, and Timeline.

APP CAPTUREThe English light-theme Creator Relationship timeline combining messages, Samples, Campaigns, and work history.

Relationship history provides evidence; it is not a sixth queue. After making the decision or completing the manual action, return to the Workbench and update the current item so it does not remain pending indefinitely.

#Recommended daily order

1. In Agent Workbench, handle Pending Escalations that block Samples, fulfillment, or collaboration commitments.
2. Complete Pending Proposals under Pending review that affect real operations.
3. In Manual Workbench, clear Sample Applications and Creator messages.
4. Use All Agent work to investigate anomalies or verify AI work history.
5. Feed recurring decisions back into configuration so the Agent-to-human boundary becomes clearer.