The slowest part of standing up a new agent used to be the blank canvas: you know what the agent should do, but the platform wants a system prompt, a tool list, an evaluation rubric, and a channel assignment before a single conversation can happen. Devotel Orbit's from-prompt builder collapses that step — you describe the agent in plain English, and Claude generates a fully wired draft with tools, a system prompt, and evaluation rubrics pre-wired. A human reviews the draft in the Agent Hub before a single conversation goes live.
This post is the walkthrough for that archetype: when a plain-English scaffold beats the blank canvas, what the generation step produces, what it deliberately does not do, and where the review and guardrail points sit. It joins the orchestration patterns index as the practical "how a squad gets its first member" entry point.
Why a scaffold beats a blank canvas
The blank-canvas problem is not that writing an agent is hard — it is that the first draft of every field competes for the same hour. The system prompt wants to be written by whoever knows the business rules; the tool wiring wants someone who knows which capabilities the tenant has enabled; the evaluation rubric wants whoever will be answering for quality in a month. Most teams ship the system prompt and defer the other two.
A generated scaffold refills all three from the same brief:
- Every required field starts filled. The system prompt, recommended channels, tool selection, rubrics, and sample inputs arrive together, so nothing is silently deferred to a later sprint that never happens.
- The draft is a concrete starting point, not a guess. "Answer missed-call follow-ups and book callbacks" becomes a named agent with a first-pass prompt in seconds — edit a draft, don't invent from zero.
- The rubric and sample inputs exist before the first live conversation. The evaluation loop the evaluation framework describes starts with a corpus on day one, not as a retrofit after the first quality incident.
The archetype fails in one specific way: treated as an auto-ship. It is a draft-generation step — the review gate below is the feature, not an inconvenience to route around.
The walkthrough: describe, generate, review, activate
The builder lives in the Agent Hub: open Agents in your dashboard (/agents under any locale, e.g. /en/agents), click Create agent, and the dialog closes with a Build with AI link that takes you to the from-prompt page (/agents/from-prompt). The flow is four steps:
1. Describe the agent in plain English. Write what the agent should do as if briefing a new hire: the task, the tone, what it may do, and where the line is. Twenty characters minimum; the counter runs to 6,000. Good briefs name the outcome ("resolve the shipping-status question or escalate with a reference number") and the boundary ("never quote a refund amount — route refunds to the team inbox").
2. Generate. The builder streams the draft field-by-field as Claude writes it, so you can read the system prompt while the rubrics are still arriving. Generation produces a draft agent — nothing is live, and no channel starts answering.
3. Review in the Agent Hub. The draft opens like any other agent: you edit the system prompt, swap tools in or out, and check the rubrics. The Try sample inputs button fetches the seeded example inputs the generator saved with the draft, so you can see what real evaluation inputs will look like before activating.
4. Activate. Activation flips the draft to a live agent and lands on its detail page. Until you click it, the draft answers nothing.
What the draft contains — and what still needs a human
Every generated draft carries the same six artifacts, and the split between "pre-wired" and "review-required" is deliberate:
- Name and description. Sensible defaults from your brief. Review for naming conventions — the generator has never seen your org's agent taxonomy.
- System prompt. A complete first draft in a monospace field, structured from the outcome and boundaries your brief named. This is the highest-value artifact and the one that most deserves a careful read: the model writes from your words, not from policy documents it has never seen.
- Recommended tools. Tool suggestions matched to the brief — but only from the capabilities your tenant has enabled. Tools that are not in your tenant's registry are dropped rather than wired, and the draft says so ("not in registry — dropped") instead of silently shipping a thinner agent. Verify the list against what the agent actually needs to do; a missing tool you do have access to is a one-click add.
- Recommended channels. Where the agent should answer — voice, SMS, WhatsApp, inbox. Generation recommends; activation is when a channel actually starts serving traffic.
- Evaluation rubrics. Name-and-criteria rubrics seeded from the brief, ready to score conversations against. This is the piece the blank canvas most often skips, and it is the entire basis of the evaluation framework: a rubric set that exists before production is the difference between a framework and a dashboard. Read every criterion; tune the fail-conditions that matter to your compliance surface.
- Seeded sample inputs. Example inputs persisted with the draft that become the first rows of the evaluation corpus. Preview them with Try sample inputs before activating.
Where guardrails fit: the generated draft inherits your tenant's safety configuration, not a generated one. The five built-in scanners and any rules you authored with the Custom Guardrail DSL apply to a prompt-built agent exactly as they do to a hand-built one — the builder never drafts around them. Tool approvals stay where they already live: any tool that requires operator sign-off before it executes (an outbound payment, a booking write) keeps that approval gate on the draft exactly as on a manually configured agent. Compliance controls — quiet hours, consent checks, disclosure lines — remain tenant-owned configuration layered underneath, not something the generator invents.
Where it lives, and where it goes next
Open the Agent Hub at /agents → Create agent → Build with AI, which lands on the from-prompt page (/{locale}/agents/from-prompt). The generated artifact is a normal draft agent: it versions, it evaluates, it appears in the same hub list, and it is scored by the same Evaluations surface as every other agent. The generation step is how it was born; nothing about its lifecycle is special after that.
For the next step after a single scaffolded agent: the orchestration patterns index covers how drafted agents compose into squads, and the evaluation framework is the loop that turns the seeded rubrics into a promotion gate. Describe, generate, review — then let the evaluation layer decide what the agent earns next.