Agent Studio Overview
The Agent Studio is the part of Omniflow where you design, configure, version, test, and observe every AI agent in your workspace. An agent is the unit of automation in Omniflow — it’s the thing that holds a conversation with a customer, calls tools, retrieves knowledge, and decides when to hand off to a human. The Studio is where it lives across its whole lifecycle: from a draft someone sketches on a Tuesday to the production agent handling thousands of conversations a day.
What an “agent” is in Omniflow
An agent is the bundle of decisions that determine how a particular conversation goes. Concretely, an agent owns:
- A model — which LLM (and which voice model, for realtime). Swappable.
- A prompt — system instructions, tone, persona, refusal rules. Versioned.
- A knowledge attachment — which documents and FAQs it can retrieve from.
- A tool list — which actions it can call (internal Omniflow actions or your webhooks).
- A routing surface — which channels and which conversations it picks up.
- A reliability config — model fallbacks, latency budgets, transfer rules.
Two agents in the same workspace can share none of those — a “Sales SDR” agent and a “Tier 1 Support” agent live side by side with completely different configs.
The agent lifecycle
Every agent moves through these phases. The Studio shows you which phase you’re in.
| Phase | What happens | Who’s involved |
|---|---|---|
| Draft | Prompt and config exist; no traffic flows. | The author. |
| Tested | Test calls and scenario runs pass; QA sample looks clean. | Author plus a reviewer. |
| Published | The agent is bound to a channel/route and handling production traffic. | Workspace admin signs off. |
| Monitored | Live traces, scorecards, and reliability alerts feed back into the next version. | Whoever owns the agent. |
You don’t progress an agent by clicking a button — you progress it by having something to show at each phase. The Studio surfaces what’s missing.
What’s in the Studio
| Panel | What it does |
|---|---|
| Agents list | All agents in the workspace, with their model, voice, status, and last-used timestamp. |
| Configuration | The system prompt, tone, language, model, voice, and tool bindings for the selected agent. |
| Knowledge | Articles, FAQs, and documents the agent can retrieve from. |
| Tools | Functions the agent can call — internal Omniflow actions or your own webhooks. |
| Activity | Per-call traces showing every model turn, tool call, and transfer. |
| Versions | A full history of every config change, with one-click rollback. |
| Reliability | Model fallback chain, latency budgets, transfer-to-human rules. |
Standard vs Realtime
Every agent has a mode that determines how it talks:
- Standard — text-first, with optional speech-to-text and text-to-speech bolted on. Use this when you want maximum tool flexibility, retrieval, or chat-only deployments.
- Realtime voice — uses a realtime voice model for sub-300ms voice latency. Use this for inbound and outbound phone calls where the agent is the primary speaker.
You can switch modes on an existing agent, but the prompt usually needs tightening — voice prompts are shorter and more conversational.
Realtime agents stream audio bidirectionally and are subject to per-minute pricing on the underlying model. Standard agents are charged per token plus any STT/TTS minutes you run separately.
Versioning and rollback
Every save creates a new version. Versions are immutable and queryable, so you can:
- Compare two versions side-by-side.
- Roll back instantly if a change regresses behavior.
- Audit who changed what, when, and why.
First 10 minutes — building your first agent
Pick a starting template
The Studio ships with templates for Tier 1 Support, Sales SDR, Booking & Scheduling, and Knowledge Q&A. Pick the closest match — you’ll edit the prompt anyway. See Creating an Agent.
Tighten the prompt and tone
Edit the system prompt and tone profile to your brand. Don’t try to make the agent do everything in one pass — start narrow. See Prompts & Tone.
Attach knowledge
Connect at least one knowledge source — a help center URL, a Notion space, or uploaded PDFs — so the agent can retrieve real answers instead of guessing. See Knowledge Sources.
Add the tools it actually needs
Resist the urge to add every tool. Add the two or three actions the agent will call most often. See Tools & Actions.
Run a test conversation
Use the Studio’s test panel to run a few scripted conversations. Read the trace — every model turn and tool call is logged. See Activity Logs & Traces.
Configure reliability
Set the model fallback chain and the transfer-to-human rules before you publish. See Reliability & Failover.
Publish to a small surface first
Bind the agent to one channel or one routing rule before opening the floodgates. Watch the first day of traces.
Where to go next
Related
| If you want to… | Go to |
|---|---|
| Build your first agent end-to-end | Your First Agent |
| Understand the voice runtime | Voice Runtime |
| Train humans alongside AI agents | Training Overview |
| Score agents at scale | QA Analytics Overview |