Warm up a new agent (resolve its setup tasks)
2 min · 6 steps
A freshly built agent shows a warmup card listing everything it needs configured. Required steps must be done before the first run; recommended ones are optional.
The warmup card on Overview lists the placeholders the planner couldn't resolve — required steps (which BLOCK the first run) sorted first, then recommended polish. They cover: missing OAuth connections (Microsoft / Google / Slack / …), missing API-key connectors (SMTP, Resend, Twilio), and inline placeholders in node configs (URLs, prompts, file paths). 'Run now' stays disabled until the required ones are cleared — resolving each saves straight into the agent's node settings.
Steps
- Work the required steps top-down.
Each task knows which surface fixes it. OAuth tasks open a focused popup right on the Overview; connector tasks open the unified picker; builder tasks jump to the node in the Builder. Resolved tasks drop off the list.
- Pick an existing connection or add a new one.
If you've already connected the provider for another agent, the picker shows your connection at the top — one click binds it.
- Warmup auto-runs once everything required is done.
When a freshly-built draft agent has no required setup tasks left, the warmup starts on its own — it dry-runs the agent and auto-repairs any node that trips, showing each step (which node failed, what the fix was, the re-run) as a live timeline. You can still trigger it manually with 'Run warmup', and 'Run anyway' bypasses the gate (the run still validates server-side).
- Warmup escalates to your cloud model when the local one is stuck.
If the local model can't fix a failing step after a couple of tries, warmup escalates to your configured cloud model rather than dead-ending. In cloud-only routing this happens automatically. In local-first mode (the default) it ASKS first — warmup pauses and shows an 'Escalate to cloud' button so your agent's config only leaves your machine when you click it. (If no cloud model is configured, you're pointed at Settings → AI Model for Build to set one.) Every escalation is recorded in the Activity stream so you can see when, and to which model, a fix climbed.
- When warmup needs your input.
If the warmup hits something only you can supply — a local file path the agent's filesystem scope must allow, a website its guardrails don't yet allow, a mailbox, a verified sender address, a value it can't invent — it pauses and shows a 'Provide input' button. Click it to pop the right surface: a file/folder picker for paths, an 'Allow web access' prompt when a fetch is blocked by your allowed-domains guardrail (allow just that site, or switch to fetch-from-anywhere — private/internal addresses stay blocked either way), or a focused question box for other values (it writes your answer into the right node — no hunting in the Builder). What you enter is saved into the agent for every future run, then warmup re-runs to confirm. A small 'Open Builder instead' link is there if you'd rather set it by hand. For a value that genuinely changes every run, use a 'User input' node in the workflow instead.
- If warmup still can't finish a step.
After trying the local model, escalating to cloud, and exhausting its repair budget, warmup stops on that one step with a clear reason and an 'Open this step in the Builder' button that jumps straight to the failing node — every repair it DID land along the way is already saved. No more being dropped on a blank editor with nowhere to go.
Live recipes need the desktop
This article is a static preview. The in-app Help sidecar inside Avery NXR can fire each step against your live project — install the desktop to use it interactively.