Appearance
Agno Workflow Builder
The Agno Workflow Builder is a second, newer workflow builder. It builds agent pipelines: a chain of steps, most of which are an LLM agent with a model, some instructions and a set of tools. Workflows are saved to the database, run from the panel, and can be put on a cron schedule.
Recorded recipes run from the Recorder tab, where you choose fresh input bindings and review each run. They cannot be scheduled. Their schedule dialog explains this and lets you disable a legacy schedule if one is still saved.
It is unrelated to Jarvis. The two do not share workflows, storage, step types, or the saved-workflow list.
Which builder do I want?
| Agno Workflow Builder | Jarvis | |
|---|---|---|
| Where | Panel, inside a space | Full page at /workflows |
| Built around | Agent steps with tools | Nodes that drive Mantis panels |
| Mantis access | Through the Mantis MCP tool source on an agent | Through dedicated Mantis Control nodes |
| External calls | HTTP Request step, Composio and custom MCP servers | Custom Python node |
| Scheduling | Per workflow, from the Workflows tab | Per workflow, from the Workflow menu |
| Triggers | Cron and manual run | Cron, manual run, and a webhook endpoint |
Reach for Agno when the work is "run these agents in order, with these tools". Reach for Jarvis when the work is "open this panel, make that bag, run this notebook".
Where to find it
View > Agents > Agno Workflow Builder.
The panel has three tabs: builder, workflows, and recorder.
The builder tab
Three columns under a toolbar:
- Steps on the left: the palette.
- The canvas in the middle. With nothing on it, it reads "No steps yet / Drag steps from the left panel".
- Configure on the right, which fills in when a step is selected.
Toolbar
The workflow's name is an editable field on the left. On the right:
- Open: a dropdown listing your saved workflows with their last-updated time, plus New workflow, a refresh button, and Load more when there are more pages.
- Save: saves the workflow, briefly showing "Saved".
- A trash button, shown once the workflow has been saved.
- An Input... box, whose value is passed into the run. Press Enter to run.
- Run, which becomes Stop while a run is in progress.
While a workflow is running, a strip below the toolbar streams tokens live, then shows the final result or the error.
Running always saves first, so Run on a new workflow creates it.
Steps
Drag a step from the palette onto the canvas.
| Step | What it does |
|---|---|
| Agent Step | Run a single AI agent |
| Parallel | Run steps concurrently |
| Loop | Repeat until a condition is met |
| Condition | Branch on a condition |
| Input Map | Route a field from any step into the next one |
| For Each | Run an agent once per item in a list |
| Sub-Workflow | Run another saved workflow |
| HTTP Request | Call an external HTTPS API |
Parallel, Loop and Condition are containers; the rest are leaves.
Configuring a step
Select a step to fill the Configure column. Every step has a Name. Beyond that the fields depend on the kind.
Agent Step and For Each
- Agent Config: Agent Name, Model (defaults to
gpt-4o-mini), and Instructions. - Tools: three sources, each independently switchable, and each with a searchable checklist so you can include only the tools you want:
- Composio: third-party integrations such as GitHub and Gmail.
- Mantis MCP: access to your active Mantis space. This is how an Agno agent reads and drives Mantis.
- Custom MCP Servers: add any number of HTTPS MCP endpoints by URL.
For Each additionally takes a Source Step, a List Field to iterate over (leave it empty if the step's output is already a list), and an execution setting.
Loop, Condition, Sub-Workflow
Loop takes Max Iterations. Condition builds a rule list with an expression preview under it. Sub-Workflow gives you a searchable picker over your saved workflows.
Input Map
Source Step (empty means the previous step) and Field (empty passes the whole output through).
HTTP Request
Method, URL, an optional raw body, a timeout in seconds, and toggles including Follow redirects. The URL must be HTTPS and supports placeholders: {{input}}, {{previous_step_content}} and {{steps.name.content.field}}.
Pin Output
Leaf steps get a Pin Output control. Run the workflow once, then pin a step to reuse the output it produced instead of re-running it. The pinned value is shown in an amber block; unpin to let the step run normally again. This is for iterating on the back half of a workflow without paying for the front half every time.
Delete step sits at the bottom of the Configure column.
Building and running
- Drag an Agent Step onto the canvas.
- Configure it. Give the agent a name, a model, and instructions. Switch on Mantis MCP if it needs to see your space.
- Add and connect further steps for the rest of the pipeline.
- Type into the Input box if the first step expects something.
- Click Run. The workflow saves, then output streams into the strip under the toolbar.
The workflows tab
A card grid of your saved Agno workflows. Each card shows the description, a schedule badge when one is set, and when it last ran. Card buttons:
- Run now.
- Schedule.
- Edit in Builder, which loads the workflow into the builder tab.
- Delete, which asks for confirmation first.
Scheduling
The schedule modal is titled Schedule Workflow:
- Enable schedule: "Automatically run this workflow on a cron schedule".
- Cron expression: free text, with presets for every minute, every hour, daily at 9:00, every Monday, and the 1st of the month. The format hint reads
minute hour day month weekday. - Space state: which space state the run should use for its Mantis MCP tools. If no space states have loaded, the modal says so. Leave it as "— None —" for a workflow that does not touch Mantis.
- Default run input: optional text sent to the workflow on each scheduled run.
Limits
- No webhook trigger. Scheduled and manual runs only.
- No execution history panel; the Workflows tab shows only when a workflow last ran.
- Stopping a run closes the panel's connection to it and returns the UI to idle. It does not confirm that the backend stopped.
- Pins are saved with the workflow, so a pinned step stays pinned the next time you load it. Unpin before a real run.
Record a native collection workflow
The recorder tab captures real Mantis operations inside this same movable workflow panel. It does not turn a typed description or a submitted trace into proof that an action happened.
- Choose a current authorized map and its collection, then begin a recording. Select current source pages and original embedding anchors. Incomplete source coverage blocks capture rather than silently selecting a sample.
- Supply exactly two distinct current-page demonstrations from one source host, with each column's value and complete original labelled line, then preview extraction. Review the current preview before extracting. The resulting rows remain pending: open Dataset and use its complete source-evidence review to accept them.
- Refresh the recording, then choose a current saved lens revision. If needed, link a current accepted record to a selected original anchor using an exact collected-evidence quote, and separately review those links. Existing currently grounded links remain on the server. The lens ranks linked original embeddings; records inherit their linked original's ordering. This does not create record embeddings or establish a scientific claim. Rejected or pending records are not treated as accepted.
- Review the accepted ranked subset and destination dashboard, name the new map, and publish. Refresh processing or uncertain publication receipts until they settle. Failed or stale publications cannot complete the captured chain; keep their receipts and stop capture once all work is terminal.
- Compile the genuine completed four-operation chain. Open Builder to edit the surrounding branches and checks, then save changes.
Recorded native nodes keep their source-reference types and recording identity through save and reload. Their old cached outputs cannot be pinned or reused as acceptance. The ordinary toolbar Run asks you to use the Recorder replay flow for recorded recipes; ordinary agent workflows retain their existing run behavior. The recorded-recipe palette offers Condition, Parallel and Input Map; native actions come from the captured chain. Keep every native source connection visible when editing the canvas.
Replay with fresh bindings
Choose Bind a second collection. Pick that collection's current pages, original anchors, new extraction demonstrations, saved lens revision and map destination. Review these bindings and prepare a fresh replay. Previously selected source IDs and row acceptances are not copied into the recipe.
Start the prepared replay and inspect its actual receipts. Extraction first pauses for preview review. After applying that preview, use Dataset to accept or reject the newly pending rows, refresh the current receipt, and supply any fresh evidence links before its second approval. Old record IDs and links are never replay bindings. Each map review pauses on its current evidence and basis. Editing a destination, changing a lens revision or receiving a different review basis retires the related consent. Resume continues the same run ledger.
When a finished map returns the replay to ready, use Start reviewed replay to continue that same ledger through its remaining saved steps and branches. The replay reports completion after the workflow finishes its actual execution path.
If a worker stops before finishing, refresh the run's receipts. When the platform can confirm that dispatch has ended, review the Interrupted replay notice and check its current-receipt confirmation before resuming. A different interrupted dispatch requires fresh confirmation. Pending native reviews and uncertain map publications still need their existing recovery; this confirmation cannot approve them. If worker completion cannot be verified, the run remains pending.
If an acknowledgement is lost, the panel retains the original request ID and body in this browser session. Close/reopen or reload, then Retry unchanged. A failed preflight during that retry does not erase the earlier uncertain action. Changed commands remain locked until recovery; processing is never labeled completed. Recovery is scoped to the signed-in account and Space and is unavailable when session storage cannot safely retain the action.
Optional Browser Apprentice
For a current authorized taught procedure, the recorder shows its revision, required parameter names and current availability. Open the existing Dataset Apprentice tab to pair, teach, inspect or resume using fresh selected-tab consent. A recorder start requires fresh consent and explicit parameter values; pairing capabilities are not copied into the recipe. Waiting and paused browser receipts require attention and do not count as completion. Acquisition must finish before extraction freezes the collection.
Availability and limits
Extraction requires the deployed native Dataset Studio integration and a current saved lens. Missing dependencies are shown as unavailable; the recorder does not substitute another extractor or invent successful receipts. Options lists are capped at 50; current collection context supports at most 40 pages and 1,000 original anchors. Truncated source context blocks this first recording family. Requests are bounded to 128,000 bytes, and required procedure parameters are capped at 30.
Stop capture and begin another recording
Use Stop capturing this recording above the bindings when this recording is finished. It releases one of your twenty currently accessible active capture slots and retains its action receipts and saved recipes. Pending browser/publication actions must first finish or fail; an uncertain operation is never discarded to release a slot. Closed recordings remain inspectable, compilable and replayable with current access and fresh review. They reject new capture. Start another recording clears the current panel pointers and begins a separate reviewed recording. Recordings whose space, origin map or collection is no longer accessible do not consume this usable quota. Their access restrictions and pending native receipts remain intact; quota recovery does not close or abandon their work.
A lost stop acknowledgement keeps the same close request for Retry unchanged. Retrying the original begin acknowledgement still returns its now-closed identity; it cannot silently reopen capture.
Native walkthrough and evidence
The recording uses invented records and original model vectors with actual native Recorder/Dataset/Builder components, authenticated Django/CSRF/PostgreSQL, original native pipeline writers and durable SDK storage. Provider inference is a disclosed input-derived local fixture; selected-tab acquisition, OAuth and deployment are not claimed as live external services. It retains its frozen earlier source composition rather than being relabelled after subsequent main integration. The same timeline is exported silently at four frames per second for compact review. Current actual-main backend revalidation and the separate capture-stop HTTP/database checks are documented in the source evidence.