Running 10+ voice agents? Ship changes without breaking live calls. Start with a free fleet review.
We help teams running 10 or more voice agents on Vapi, Retell or ElevenLabs change prompts safely: versioned config, tested rollouts and rollback. A free 30-minute fleet review call comes first.
- Draftv8 draft
- Test gatescenarios pass
- Publishv8 live
- Phone numberpoints at v8
Why do voice agents break when you change a prompt?
Because a change can go live before anyone has tested it, and because one edit can reach many agents at once. Nobody sees the damage until a caller does.
- Publishing is live
- The Vapi docs say a published version takes live calls right away, before you rerun your simulations.
- Pointers go stale
- Retell staff on the Retell community forum say a phone number pinned to a specific version does not follow new versions.
- Shared parts reach many agents
- A Retell library subflow edit reaches every agent that uses it, published versions included. Vapi tools follow the latest version unless pinned. An ElevenLabs tool edit applies to every agent and branch that uses the tool.
What is the safe way to change 10 or more voice agents?
Keep every agent's config in one versioned place, change it once, test it against saved call scenarios, roll it out in steps and keep a way back. The six steps below are the whole method.
Take an inventory
List every agent, the platform it runs on, the phone numbers that point at it and the shared tools it uses. You cannot change a fleet you cannot list.
Keep config in versioned files
Store each prompt and config as a file in git, next to the platform's own version history. Git shows who changed what across all agents, not just inside one dashboard.
Roll out one change to many agents
Write the change once, apply it to a small group, then to the rest. Read every agent back to prove the change landed.
Put a test gate before publish
Run saved call scenarios against the change. Publish only when the pass rule you wrote down is met.
Plan the rollback before the change
Save what each agent and phone number pointed at. If calls go wrong, move the pointer back first and look for the cause second.
Keep staging apart from production
Try changes where real callers cannot hear them, then promote the same files. The Vapi docs suggest one org per environment, the Retell docs suggest separate workspaces, and ElevenLabs uses branches.
What is native on Vapi, Retell and ElevenLabs, and what do you script?
All three keep versions and ship some testing. None of their docs describe a bulk update endpoint, so rolling one change out to many agents is scripting work.
Vapi
- Native
- Draft, publish, history and restore for assistants and tools (versioning shipped in the week of 17 August 2026), traffic split by percentage, Simulations, Evals and Scorecards.
- You script
- Updating many assistants, because the docs list no bulk endpoint. Reading each one back. Saving snapshots you can restore from.
Retell
- Native
- Numbered versions with drafts, tags such as prod and staging, a traffic split on the phone number, simulation tests and batch tests.
- You script
- Updating and publishing each agent, moving tags and number bindings, and keeping your own copy of prompts and LLM config.
ElevenLabs
- Native
- Immutable versions on branches, traffic deployments by percentage, simulation, next-reply and tool-call tests, and a CLI that pushes config files.
- You script
- Reading every agent back after a push, sending traffic back to Main when needed, and checking which agents use a shared tool before you edit it.
Checked against the Vapi, Retell and ElevenLabs docs on 18 November 2026.
Where should I start reading?
Start with the guide that matches what hurts most today. Each one answers a single question and lists the official pages it relies on.
- Published, but calls still use the old prompt
How each platform decides which version answers a call, and a five-step check.
- Roll out one change to many agents
No bulk endpoint, shared parts as hidden blast radius, and a read-back step.
- Versioning and rollback
What is native on each platform, what you script, and the Retell deleted LLM trap.
- Regression test after a prompt change
A test gate built on each platform's own tests, and when a dedicated tool helps.
What do we do, and what do we not do?
We set up the release process and the scripts around your own platform accounts. We are an integrator, so there is no product to sign up for.
We do
- Run a free 30-minute fleet review call: how your agents are versioned today, where a change can go wrong, what to fix first.
- Quote a fixed-scope setup after the call: versioned config, a rollout script with read-back, a test gate and a rollback runbook.
- Run the process for you afterwards, if you want that.
We do not
- Build voice agents from scratch, tune voices or pick a platform for you.
- Ask for data or access before the call. Access comes after the call and a signed NDA, with keys you create and revoke.
- Sell a console, dashboard or other software.
What do teams ask about changing many voice agents?
These are phrased the way operators ask them on the vendor forums. The short answers are below, with the detail in the guides.
I published a new prompt, so why do my live calls still use the old one?
On all three platforms the phone number or traffic setting decides which version answers. A Retell number can be pinned to an older version, a Vapi traffic split pins its versions, and an ElevenLabs number can name a branch. Read the pointer before you edit the prompt again.
How do I update the same prompt change across 30 Vapi assistants?
The Vapi docs list no bulk update endpoint. You list the assistants and update each one, or keep them as files in the open-source Vapi GitOps repo and apply from there. Either way, read each assistant back afterwards to confirm the change landed.
Can I roll back a Retell agent after a bad change?
Yes, by version. Create a draft from any past version, or move the number or tag back to the earlier version. A deleted Retell LLM cannot be recovered, according to Retell staff on the Retell community forum, so keep your prompts in your own version control.
How do I regression test a voice agent after a prompt change?
Save real call scenarios as tests and run them before you publish. Vapi has Simulations and Evals, Retell has simulation and batch tests, and ElevenLabs has simulation, next-reply and tool-call tests. Decide in advance which pass rate blocks the change.
What happens to my agents when I edit a shared tool or subflow?
It can change many agents at once. Vapi tools follow the latest version unless pinned, Retell library subflows reach every agent that uses them, and ElevenLabs applies a tool edit to every agent and branch that uses it. List the agents that use a shared part before you edit it.
Can I move agents from staging to production, or between workspaces?
It depends on the platform. The Vapi docs suggest one org per environment and a GitOps promote step. Retell staff on the Retell community forum say to export the agent JSON and import it into the other workspace. In ElevenLabs, environments are for environment variables, not for separate agent copies.
Want someone to review how your fleet ships changes?
The free 30-minute fleet review call covers how your agents are versioned, how a change reaches live calls and how you would roll it back. We ask for no data and no access before the call.