How Do I Have One AI Agent Check Another Agent's Work Before the Workflow Continues?

Summary

The steps

1. 1

   Confirm roles and create the reviewing agent

   In Dave by voolama, an Agent Review node requires two prerequisites: a user account with the right role, and a second agent whose job is to review rather than produce output. These are separate requirements and both must be in place before the node will work.

   Dave by voolama has six default roles: Admin, Create, Curate, Use, Reporting, and Financial. All six are editable. The Workflow Editor is accessible only to users with the Admin or Curate role. Creating and editing agents requires the Admin or Create role. A user with the Curate role can place an Agent Review node on the canvas and assign an existing agent to it, but cannot create or edit the reviewing agent itself. If you cannot see the Agents section or the Workflow Editor in the sidebar, your role assignment is the first thing to check. The resource at Why Can't I See Agents, Workflows, or Other Sections in Dave by voolama? explains the role-to-section mapping in detail.

   Navigate to Agents in the sidebar. Click New Agent in the top-right corner. Name the agent to describe its reviewing function, for example 'Fact Accuracy Reviewer' or 'Draft Quality Checker.' Add an optional description and tags. Click Create Agent. The agent appears in the list with zero versions. An agent with zero versions cannot be assigned to a workflow node, so you must create at least one version before the node will work.

   Do this

   • Confirm your user account has the Admin or Create role (to create the agent) and the Admin or Curate role (to edit the workflow). If not, ask your Admin to update your role under Admin, Settings, Users and Roles.
   • Navigate to Agents in the sidebar. If the link is not visible, see Why Can't I See Agents, Workflows, or Other Sections in Dave by voolama?
   • Click New Agent. Enter a name that describes the reviewing function. Click Create Agent.
   • The agent detail page opens. The Versions section is empty. Proceed to Step 2 to create the first version.

   Example

   A hypothetical small marketing agency might name the reviewing agent 'Brand Voice Checker.' The drafting agent writes social media captions. The reviewing agent checks each caption against a list of brand rules: no superlatives, no price claims, no competitor mentions. Two separate agents, two separate system prompts, one workflow.

   Best practice

   • Give the reviewing agent a name that is clearly distinct from the drafting agent. When assigning agents to nodes in the Workflow Editor, both agents appear in the same dropdown. A clear name prevents assigning the wrong agent to the wrong node.
   • Do not reuse the drafting agent for the reviewing step. The reviewing agent needs a different system prompt with different instructions. Create a separate agent.
2. 2

   Write the reviewing agent's system prompt and Config JSON

   The reviewing agent's behavior is defined entirely by its Config JSON. The five fields the Dave by voolama help documentation describes are: systemPrompt (the instruction string), model (the specific model, for example gpt-4o or claude-3-sonnet), temperature (0.0 is fully deterministic, 1.0 is most varied), maxTokens (the maximum response length in tokens), and topP (nucleus sampling parameter). For a reviewing agent, the two most important are systemPrompt and temperature.

   Why temperature matters for review agents. Temperature controls how much the model varies its output for the same input. At 0.1, the reviewing agent produces the same verdict for the same draft on every run. At 0.7 or higher, the same draft might pass one run and fail the next, which makes the review unreliable. Set temperature to 0.1 for any agent whose job is to check rather than create.

   On the agent detail page, click Create Version in the Versions section header. A form panel expands with a monospaced Config JSON editor. Enter a valid JSON object. Click Create Version. The system assigns version number v1 automatically. Version numbers are auto-incrementing integers (v1, v2, v3). Each version is immutable once created. If you need to change the system prompt later, create a new version rather than editing v1.

   Do this

   • On the reviewing agent's detail page, click Create Version.
   • In the Config JSON editor, enter a JSON object with at minimum: systemPrompt, model, temperature (set to 0.1), and maxTokens.
   • In the systemPrompt, name the exact criteria the agent should check. Instruct it to return a clear pass or fail verdict and to list any specific problems it found. Example structure: {"systemPrompt": "You are a quality reviewer. Check the draft against these rules: [list your rules here]. Return PASS if all rules are satisfied. Return FAIL followed by a numbered list of the specific rules that were violated. Do not suggest rewrites.", "model": "gpt-4o", "temperature": 0.1, "maxTokens": 512}
   • Click Create Version. The version is assigned v1 and is now available for assignment in the Workflow Editor.

   Example

   A hypothetical small manufacturer checking product descriptions might write a system prompt that lists five rules: no unverifiable performance claims, no competitor comparisons, no pricing, model number must appear in the first sentence, and the description must not exceed 150 words. The agent returns PASS or FAIL with a numbered list of violated rules. The reviewer sees a structured verdict in the Task Inbox Payload rather than a free-form AI opinion. This is the same pattern the product descriptions guide at What Is the Safest Way to Use AI for Product Descriptions Without Publishing Invented Facts? uses for its accuracy check step.

   Best practice

   • Set temperature to 0.1 for a reviewing agent. Review is a deterministic task: you want the same draft to produce the same verdict every time. A higher temperature introduces variation that makes the review less reliable.
   • Instruct the agent to return a structured verdict (PASS or FAIL followed by specifics) rather than a narrative judgment. Structured output is easier to read in the Task Inbox Payload section and easier to route on in a Routing node.
   • Keep the system prompt focused on a single category of check. If you need to check for factual accuracy AND brand voice AND word count, consider whether three separate reviewing agents would be clearer than one agent trying to do all three.
3. 3

   Place the Agent Review node on the workflow canvas

   Navigate to Workflows in the sidebar and open the workflow you want to add the review step to. Open it in the Workflow Editor. The editor uses a three-panel layout: the Node Palette on the left lists all ten node types, the Canvas in the center is where you place and connect nodes, and the Properties Panel on the right appears when you select a node.

   Find Agent Review in the Node Palette. Its code is agent_review and its icon is pink with a bot symbol. Drag it onto the canvas and position it after the node whose output you want to review, typically an Agent Interaction node. Connect the nodes by dragging from the source handle (bottom circle) of the upstream node to the target handle (top circle) of the Agent Review node.

   Rename the node in the Properties Panel to something descriptive, for example 'Check Draft Accuracy' or 'Brand Voice Review.' The default label 'Agent Review' tells a future editor nothing about what this specific node checks.

   The blog draft and review example at Example Workflow: Blog Draft and Review shows how an Agent Review node fits into a complete workflow with a drafting step before it and an editorial approval step after it.

   Do this

   • Open the Workflow Editor for your workflow. If you do not see the Workflow Editor link, confirm your role is Admin or Curate.
   • Locate Agent Review in the Node Palette on the left. Its icon is pink with a bot symbol (code: agent_review).
   • Drag the Agent Review node onto the canvas. Position it below the Agent Interaction node whose output it will review.
   • Connect the Agent Interaction node to the Agent Review node: hover over the source handle (bottom circle) of the Agent Interaction node, drag to the target handle (top circle) of the Agent Review node, and release.
   • Click the Agent Review node to select it. In the Properties Panel on the right, change the Label field to a descriptive name.

   Example

   A hypothetical agency workflow has three nodes before the Agent Review: Start, User Interaction (collect the content brief), and Agent Interaction (draft the social caption). The Agent Review node is placed after Agent Interaction and labeled 'Check Caption Against Brand Rules.' The connection runs from the source handle (bottom) of Agent Interaction to the target handle (top) of Agent Review.

   Best practice

   • Place the Agent Review node immediately after the node whose output it checks. Do not insert other nodes between the drafter and the reviewer unless those nodes transform the output intentionally.
   • Name the node after what it checks, not after the node type. 'Check Caption Against Brand Rules' is more useful than 'Agent Review 2.'
   • Save frequently. Each click of Save Version creates a new immutable snapshot. The Workflow Editor does not auto-save.
4. 4

   Assign the reviewing agent in the Properties Panel

   With the Agent Review node selected on the canvas, the Properties Panel on the right shows an Agent Assignment section. This is where you connect the reviewing agent you created in Steps 1 and 2 to this specific node.

   Open the Agent Assignment section. Select the reviewing agent from the agent dropdown. The dropdown lists all agents in your tenant that have at least one non-retired version. If the reviewing agent does not appear, it either has no versions or all its versions have been retired. Return to the Agents section, open the agent, and confirm at least one version exists and is not retired.

   After selecting the agent, select the version you want to use. Version v1 is the first version you created. The workflow will use the exact version you select here, even if you later create a newer version of the agent. To update the workflow to use a newer agent version, open the workflow, update the Agent Assignment, and save a new workflow version.

   Do this

   • Click the Agent Review node on the canvas to select it. The Properties Panel opens on the right.
   • Find the Agent Assignment section in the Properties Panel.
   • Select the reviewing agent from the agent dropdown. If it does not appear, navigate to Agents, open the agent, and confirm at least one non-retired version exists.
   • Select the agent version to use, for example v1.
   • Click Save Version in the top-right corner of the editor to save the assignment.

   Example

   A hypothetical agency selects the 'Brand Voice Checker' agent and version v1 in the Agent Assignment section of the 'Check Caption Against Brand Rules' node. When a workflow instance reaches that node, it invokes Brand Voice Checker v1 automatically, passing the caption from the previous Agent Interaction step as input.

   Best practice

   • Always select a specific version rather than relying on a default. Knowing exactly which version is assigned makes it easier to diagnose unexpected review results later.
   • After assigning the agent, save a new workflow version immediately. The assignment is not persisted until you click Save Version.
   • If you update the reviewing agent's system prompt by creating a new version, remember to update the Agent Assignment in the workflow and save a new workflow version. Old workflow versions continue using the agent version they were saved with.
5. 5

   Configure outgoing edges for pass and fail paths

   The Agent Review node produces a result: the reviewing agent either passes or fails the output. The workflow must have outgoing edges that handle both outcomes. Without a configured fail path, the workflow may hang at the Agent Review node when the reviewing agent returns a fail result.

   The Workflow Editor's Properties Panel includes an Input / Output Ports section for most node types. Use this section to configure the outgoing edges from the Agent Review node. Connect the pass path to the next step in the workflow, for example a Human Review node or an End node. Connect the fail path to a Routing node that sends the output back to the drafting step for revision, or to a Human Review node where a person can decide what to do with failed output.

   A Routing node (code: routing, purple icon) is a logic-only node that evaluates conditions and directs execution without creating tasks or invoking agents. It is useful for separating the branching decision from the functional nodes. For example, a Routing node after Agent Review can check whether the reviewing agent returned PASS or FAIL and route accordingly.

   Do this

   • Select the Agent Review node on the canvas. In the Properties Panel, find the Input / Output Ports section.
   • Configure a pass path: connect the Agent Review node's source handle to the next node in the workflow (Human Review, End, or another step).
   • Configure a fail path: connect the Agent Review node to a Routing node or directly to a Human Review node. The fail path should not lead to an End node unless a failed review is a valid terminal outcome for your process.
   • If using a Routing node, configure its branching criteria to check the reviewing agent's output for PASS or FAIL.
   • Confirm every path leads to an End node. Dangling paths cause instances to hang at runtime.
   • Click Save Version.

   Example

   A hypothetical agency workflow routes PASS results to a Human Review node labeled 'Final Editorial Approval' and FAIL results to a Routing node labeled 'Route Failed Caption.' The Routing node sends the failed caption back to the Agent Interaction node for a revised draft. The help documentation does not describe a built-in loop counter, so the agency uses a Routing node with a field that tracks revision count to limit the number of revision cycles.

   Best practice

   • Always configure both a pass path and a fail path before saving the workflow version. A workflow with only a pass path will hang every time the reviewing agent returns a fail result.
   • Consider adding a Human Review node after a PASS result if the stakes are high enough that a person should still see the output before it proceeds. Agent Review handles consistent rule-based checking; Human Review handles judgment calls.
   • Keep the fail path short. If a failed draft requires more than two revision loops, the drafting agent's system prompt probably needs to be improved rather than the workflow made more complex.
6. 6

   Read the review task in the Task Inbox and decide when to add Human Review

   When a workflow instance reaches the Agent Review node, the reviewing agent runs automatically. The result appears in the Task Inbox as an agent_review task. Navigate to Tasks in the sidebar to open the Task Inbox. The task table shows the task type, status, parent instance, assignee, and due date. Click the task type link to open the task detail page.

   On the task detail page, the Payload section shows the JSON data passed from the workflow to the task, including the reviewing agent's output. If the reviewing agent was instructed to return a structured verdict (PASS or FAIL with a numbered list of violations), that verdict appears here. The same three decision buttons available on Human Review tasks are available on Agent Review tasks: Approve (green), Reject (red), and Request Changes (outlined). You can add a Comment to explain your decision.

   The Task Inbox is accessible to users with the Admin or Use role. Users with Create, Curate, Reporting, or Financial roles cannot access the Task Inbox.

   When to add a Human Review node after Agent Review. Agent Review is suited to rule-based checks the AI can evaluate consistently: does the draft contain only facts from the supplied list, does it stay under a word count, does it avoid banned phrases. Human Review is suited to judgment calls where accountability must rest with a named person. For high-stakes output, add a Human Review node after Agent Review so a person sees the output even when the agent passes it. The article at How to Review AI Output Without Becoming the Bottleneck covers how to structure review steps so they do not slow the workflow down.

   Do this

   • Navigate to Tasks in the sidebar. The Task Inbox opens.
   • Find the agent_review task for your workflow instance. Its status badge shows pending (amber) if it is waiting for a decision.
   • Click the task type link to open the task detail page.
   • Read the Payload section. The reviewing agent's verdict and any listed violations appear here.
   • Click Approve if the output passed and should proceed. Click Reject if the output failed and should not continue on the current path. Click Request Changes if the output needs revision before it can proceed. Add a Comment to explain your decision.
   • Decide whether to add a Human Review node after Agent Review in the workflow. If the output is customer-facing, high-stakes, or involves commitments, add a Human Review node so a person has a final gate.

   Example

   A hypothetical agency reviewer opens an agent_review task and reads the Payload: 'FAIL. 1. Caption contains a superlative which is on the banned phrases list. 2. Caption exceeds 150 characters by 12 characters.' The reviewer clicks Reject and adds a Comment: 'Revise to remove superlative and shorten to under 150 characters.' The workflow routes the failed caption back to the drafting agent for a revised attempt.

   Best practice

   • Use the Comment field on every decision, not just rejections. A brief note on an approval creates a useful record if the output is questioned later.
   • If the reviewing agent consistently passes output that a human reviewer later rejects, the system prompt needs to be more specific. Create a new agent version with tighter criteria.
   • If the reviewing agent consistently fails output that a human reviewer would pass, the criteria in the system prompt may be too strict. Review the system prompt and create a new version with adjusted rules.

Agent Review versus Human Review: which node does what

Where this guide stops