Retell shipped a copilot that simulates every change to a voice agent before it goes live. It runs the scenario you name, shows you what the agent would actually say, and never touches a production agent without your approval.
If you run AI calling campaigns, the failure mode you know is this: a prompt tweak looks fine, you push it, and you find the crack forty dials later when a prospect asks something ordinary and the agent falls apart. Conductor is aimed squarely at that.
What Retell shipped
Conductor landed in Retell's July 2, 2026 changelog. Retell describes it as its "copilot for building and maintaining voice agents." It has three parts.
Simulation before deployment. You can ask for a specific scenario in plain language, and Conductor runs the call against your agent without dialing a real number. Retell's example is "what happens if the caller wants to reschedule." You get the agent's response back in the interface. Retell says every change is validated against the platform before you see it, and frames the point as catching edge cases most teams only find after a call goes wrong.
Learning from failed calls. Conductor reads calls that went badly, writes reproducible tests from them, runs those simulations, and proposes fixes. Retell says it already handles 70% of Retell's own simulation testing internally.
Human review on every edit. Conductor never edits a live agent on its own. Proposed changes surface inside the actual agent flow with old and new values side by side, and you accept or reject one change at a time, with undo.
Two smaller items shipped in the same entry. Free LLM tokens went up to 4,000 per agent. And knowledge base parsing improved: tables and headers parse better, chunks retain their parent header information, and text splits at sentence boundaries instead of mid-word.
Why this matters for outbound teams
Voice agents have always had a testing problem. A cold email sequence can be proofread. A voice agent is a branching conversation, and the only honest way to test one has been to call yourself, call a teammate, or let it loose on a list and read transcripts afterward. All three are slow, and the third one burns real prospects to find bugs.
That cost is not theoretical. A list of 500 wholesalers or 300 property managers is a finite asset. Every contact the agent mishandles is a number you cannot dial again with a straight face. Testing on live pickups means paying for QA in leads.
The knowledge base change matters more than it reads. Most of the information a B2B agent needs (pricing tiers, service areas, qualification rules) lives in tables. Chunks that keep their parent header stop the agent from confidently reciting a price from the wrong row. Chunking mid-word is how an agent ends up quoting half a sentence back to a prospect.
The 4,000 free token bump is a testing-budget story, not a production one. It lowers the friction on building an agent to the point where it is worth demoing.
Our read: the review gate is the part that makes this usable on client accounts. An AI that proposes edits and shows you a diff inside the flow is a tool. An AI that silently rewrites a live agent handling real calls for a paying client is a liability. Retell chose correctly.
How we'd use it
We build and run outbound voice agents on Retell for clients in home services, inspections, and B2B services. Here is where Conductor fits our actual workflow.
1. Turn every real objection into a permanent test. Our agents die in the same places: "how did you get my number," "are you a robot," "just send me an email," a gatekeeper who wants a last name. Today those live in a doc and get re-tested by hand. With on-demand simulation, each one becomes a scenario you replay after every prompt edit. Build the list from your own call recordings, not from imagination, and run the whole set before any change ships.
2. Pre-flight a new client agent before it touches the list. The first 50 dials of a new campaign are the most expensive, because that is when the script is weakest and the list is freshest. Run the qualification path, the transfer path, the callback-booking path, and the not-interested path in simulation first. Reserve live dialing for problems that only surface with real humans, like timing and interruption handling.
3. Close the loop from failed calls to fixes. Conductor reading bad calls and writing reproducible tests from them is the useful half of QA that nobody does manually. Pair it with your own review: pull the calls where the agent got hung up on in under fifteen seconds, let Conductor build tests from them, then judge the proposed fixes yourself before accepting.
4. Fix the knowledge base before you blame the prompt. If your agent is quoting the wrong price or the wrong service area, the prompt is often innocent and the retrieval is at fault. Re-upload pricing and service tables now that headers and sentence boundaries are handled properly, then simulate the specific questions that were producing wrong answers.
FAQ
Can you test a Retell voice agent without making a real call?
Yes. Conductor simulates a scenario you describe against your agent and returns the response without dialing a number.
Does Conductor change my voice agent automatically?
No. Retell states Conductor never edits a live agent on its own. Changes appear in the agent flow with old and new values side by side, and you accept or reject each one individually, with undo.
Is Retell Conductor worth it for a small outbound team?
If you run more than one agent or edit prompts regularly, the value is in not burning leads to find bugs. A single-agent team calling a large list gets less out of it than an agency maintaining several client agents.
Get your AI calling system built right
AnaqVisual builds and runs outbound systems for B2B teams: cold email infrastructure, AI calling agents on Retell and Bland, and Clay-based list building and enrichment. We do this every day, which is why we care about testing gates most people skip.