Agents#
Agents connects images, models, conditions, and actions in a visual workflow. Use YOLO to decide which images need a closer look, ask a vision-language model to explain them, or collect selected images in a dataset for review.
Switch to your personal workspace using the sidebar workspace selector and enable Early access in Settings > Profile. Then open Agents in the sidebar or go to platform.ultralytics.com/agents. Early access is a personal preference, including when you work in a team workspace.
Run Your First Workflow#
- Open Agents, click Add block > Templates, and choose YOLO → VLM Monitor.
- Select the Input image block, choose an uploaded dataset, and select an image containing a person. Use an image of up to 4 megapixels for the shared trial.
- Keep the default YOLO26 nano model. The condition passes when YOLO detects at least one person.
- Select Describe the scene, choose a vision-language model, and enter a prompt. Add the matching provider key in Settings > API Keys if you have not already done so. In a team workspace, ask the workspace owner to add or replace the provider key.
- Keep Shared trial · 60 seconds selected for a small run, then click Run.
- Watch the blocks update and inspect the Output. If no person is detected, the condition finishes without running the description branch.
You can run a draft before saving it. Use Save workflow to name and save a workflow for later, and save again after making changes. The Python button shows the generated workflow code.
Start with a Template#
| Template | Workflow | Useful for |
|---|---|---|
| YOLO → VLM Monitor | Image → YOLO → condition → vision-language model → Output | Describing a scene only when a person is detected |
| Collect Uncertain Detections | Dataset → Deployment → confidence condition → Dataset | Building a review set from images with uncertain detections |
| Capture and Alert | Image → YOLO → condition → Dataset and Slack | Saving an image and notifying a channel from the same condition |
Templates are editable starting points. Select the images, deployment, destination dataset, and integrations for your workspace before running them. These workflows run when you click Run; the Monitor template analyzes the selected input image.
Choose Where to Run#
The execution selector next to Run chooses the hardware for the workflow.
| Choice | Use it for | Limits and pricing |
|---|---|---|
| Shared trial · 60 seconds | Trying a small workflow | Up to 60 seconds, 20 total input images of up to 4 megapixels each, and 20 model/provider calls; one local YOLO model, using an official YOLO26 nano model |
| Existing dedicated deployment | Larger datasets, longer runs, or other local models | Uses that deployment's existing hardware and pricing, including an eligible free dedicated deployment; no overall workflow duration or image-count cap |
| New deployment… | Creating dedicated hardware from the builder | Opens the existing deployment creation and pricing dialog; choose resources and wait until the deployment is ready |
Shared trials allow two submissions per five minutes. Platform runs support up to 20 blocks. Model/provider calls include YOLO, vision-language models, deployment predictions, and Slack actions; branching can increase the number of calls per input image.
Dedicated runs continue in checkpointed requests on the selected deployment. Closing or reloading the browser does not stop a run. Existing deployment limits and provider limits still apply, and language-model usage is billed by the provider associated with your key. See deployment pricing for hardware options.
A Deployment block selects the endpoint used for a prediction step. The execution selector chooses where the workflow itself runs. You can choose the same dedicated deployment for both.
Configure Blocks#
| Block | Purpose |
|---|---|
| Image | Select one image from an uploaded dataset |
| Dataset without an incoming connection | Read images from a selected dataset and split, with a configurable input limit |
| YOLO | Run an official model or a trained model from the workspace |
| Deployment | Predict through an existing ready Platform deployment |
| LLM | Send the image and upstream context to the selected language or vision-language model |
| Gate | Control downstream execution with a condition or cadence |
| Dataset with an incoming connection | Add the incoming image to a destination dataset |
| Slack | Send a message to the workspace's connected Slack channel |
| Export | Export a model using Platform's existing export workflow |
| Output | Display the upstream result |
Select a block to edit its settings. Connect an output to another block to pass its result downstream. One output can connect to multiple blocks: for example, connect one condition to both a destination dataset and Slack. Each branch uses the same evaluated condition.
Filter Detections with a Condition#
Use the same Gate block for detection counts and confidence filters. Choose the upstream result, any class or a specific class, and a comparison. For example:
- Person count ≥ 1: continue when at least one person is detected.
- Confidence between 0.25 and 0.5: continue when at least one matching detection falls within that inclusive range.
- A specific class in a confidence range: collect uncertain detections for that class only.
Confidence filters examine individual detections. An image with scores of 0.3 and 0.9 passes a 0.25–0.5 range because one detection matches. Images with no detections do not pass a confidence condition. Set the prediction model's confidence cutoff low enough to retain the detections you want to evaluate.
A condition that does not match completes with no downstream output. Its connected actions do not run for that image.
Collect Images for Review#
Open Collect Uncertain Detections, choose a source dataset and split, select your prediction deployment, and configure the confidence range. Select an existing destination dataset in the final Dataset block. Create a dataset from the Datasets page first if needed.
The destination receives the original images, unlabeled, in its train split. Existing copies are skipped, and normal dataset storage quotas apply. Source labels and model predictions are not copied as annotations. Open the destination in the existing annotation editor to review and label the collected images.
Use hosted images that your workspace can access and a destination dataset you can edit. Connected cloud-storage and on-premise datasets cannot be used as collection destinations.
Send Conditional Slack Alerts#
Connect Slack in Settings > Integrations, then connect a condition to a Slack block. Enter the message and use {output} to include the upstream result. Capture and Alert also connects that same condition to a dataset, so a matching image is collected and a message is sent to the connected channel.
Workflow Slack messages are configured in the block. They are separate from the training, export, and deployment notification choices in Settings.
Follow Progress and Stop a Run#
Active blocks show a green spinner. Completed blocks show a green check and green outline; failed blocks show a red outline. Counts show which inputs have reached each block, and downstream blocks remain unexecuted when a condition does not match.
Reloading the page restores the run's progress and selected execution deployment. Completed results remain visible for the matching workflow. Click Stop to cancel an active run; cancellation waits for the executing block to reach a cancellation point. Stopping or deleting its execution deployment also cancels the workflow.
Troubleshooting#
- Agents is missing from the sidebar: enable Early access in your personal Settings > Profile.
- The run exceeds the shared trial limits: select a dedicated deployment, including an eligible free deployment, or reduce the image resolution, input count, and model/provider calls.
- The deployment is preparing or stopped: wait for it to become ready or start it from Deployments.
- A provider key is missing: connect the selected provider in Settings > API Keys. In a team workspace, ask the workspace owner to add or replace the key.
- No images reach the destination: inspect the condition and prediction confidence cutoff. Empty detection results never satisfy a confidence range.
- Collected images have no labels: collection adds original images for review. Use the existing annotation tools to label them.