LumisReach
  • Pricing
Products
AI agents
Call Receptionist
Every call answered, day or night
Call Outreach Agent
Your whole list worked, not the top of it
Speed-to-Lead Agent
New leads called inside a minute
Content Marketing Strategist
Researches, films and posts every day
Conversion Rate Specialist
Turns the traffic you have into calls
Developers
API
REST for calls and numbers
MCP
Give your AI client a phone
Services
Custom Agentic Workflows
We build what no product on this list covers
Industries
E-commerce ConsultingNew
1:1 GTM expert, +50% conversion in 14 days
Wholesale Distributors
The counter line in, the whole account base out
Outbound Sales
Verify the list, work it, hand over the yeses
Dealerships
Sales floor and service drive, one number
Developers
Telephony and caller trust under your own agent
Legal Practices
Live intake, conflict fields, booked consults
Software & SaaS
Called while the signup tab is still open
Trades & Contractors
Answered on site, triaged, on the board
Freight & Logistics
Cover the load, then run the check calls
Guides
Guides
Best AI receptionist
The top picks for small business
How AI receptionists work
From the first ring to the booking
Virtual vs AI receptionist
A human service or an agent, compared
Answering service pricing
What call coverage really costs
After-hours answering service
Nights, weekends and holidays covered
Medical answering service
Call coverage for clinics and practices
Real estate answering service
Every buyer and seller call picked up
HVAC answering service
No-heat calls answered and booked
Comparisons
LumisReach vs Retell AI
Features, pricing and setup
LumisReach vs Synthflow
Features, pricing and setup
PricingSign InCreate Account
Sign InCreate Account
Back to Blog
Developers/10 min read

Can You Trust an AI Agent to Call Your Customers? Isolation, Approval and Proof (2026)

A call or a text cannot be unsent, so an AI agent with a phone line needs more than a good prompt. How running LumisReach's MCP server inside Rush (getrush.ai) splits the job: where the agent runs, how the key is held, who approves before money or messages go out, and what counts as proof the call happened.

Published November 19, 2026By LumisReach

A phone waveform sealed inside two nested boxes, representing an AI calling agent running in an isolated cloud computer

Short answer: yes, if three things are true. The agent runs somewhere isolated with only the credential the job needs; anything that spends money or reaches a new person passes a gate that a person, not the agent, answers; and "done" means a call record with an outcome and a transcript, not the agent's own summary. Running the LumisReach MCP server inside a Rush run covers most of that between the two: Rush gives each run its own microVM and keeps the agent's record, and LumisReach keeps every call and text, its outcome and its cost on the server side.

Rush details come from Rush's docs and engineering blog, checked November 19, 2026. For the setup itself, see how to give a Rush agent a phone line.

Why is a phone call riskier than a pull request?

Rush's engineering team writes about agents mostly in terms of code, and code has an undo button. A pull request can be closed and a deploy can be rolled back. A voicemail to a customer, or a text to the wrong number, cannot.

Rush's post Every Agent Demo Hides the Same Number puts this in one formula, V = min(P, R) × Q − (1 − Q) × F: the value of an agent is its production rate capped by your review rate, times quality, minus the failure rate times F, the cost of a failure. A demo sets F to zero. Outbound calls and texts are where F is highest, because the output reaches a person outside your company the moment it happens. The same post lists the property that matters most here: "Output that can be blocked. Before agent output reaches a downstream system, a review gate holds it."

The companion post, Agents Are Eating the Org Chart, draws the practical rule: "verification depth must be proportional to the cost of failure." A reminder text to someone who booked yesterday is low F. A first call to a list of 300 new leads is not.

Layer 1: where the agent runs and how it holds the key

On Rush, every run gets its own microVM with its own disk, memory, desktop and browser, shared with no other run, and each org fails closed: another org's paths answer 404 (Isolation, plainly). A run started with an API key gets only the connections it names, never your other sign-ins.

The key is the part to get right. Rush's post Your Sandbox Is Only as Isolated as Its Package Proxy walks through an incident where agents escaped a sandbox through credentials that were simply lying around in the environment, and ends with the rule: "Hand a run only the credential it needs, scoped to that run, expiring when it ends." For a phone-calling agent that means:

  • Use a dedicated LumisReach key for the Rush connection, not the one your backend uses, so you can revoke it without touching anything else.
  • Narrow it when you can. LumisReach keys can be issued with a scope list, such as voice:write and sms:write without phone_numbers:write, so the agent can call but cannot buy numbers. Keys made in Settings are unrestricted; ask LumisReach support for a scoped one.
  • Pick the tenant scope. A key tied to one location (a sub-entity) cannot read or call from a sibling location's numbers or contacts.
  • Keep it out of the transcript. Rush stores a connection token encrypted and never returns it. If your agent calls the LumisReach REST API directly rather than through MCP, save the key as a Rush header secret for api.lumisreach.com: Rush adds the header to the run's requests outside the sandbox, so the agent never holds the key at all.

Rush is honest that the sandbox is "the boundary around the run, not around the agent's tools": anything placed inside it, the agent can read. The other half of the boundary sits on our side. The LumisReach v1 layer re-checks every MCP tool call's key, scopes, rate limits and tenant on the server, so nothing the agent does inside its computer can widen what the key may do. Rush's post The Agent Turned Off Its Own Sandbox explains why that placement matters: "Nothing inside the sandbox sets what the sandbox allows."

Layer 2: who says yes before money or messages go out

Meta's consumer agent, Muse, runs a separate Sentinel agent that must approve anything leaving the machine, and "checks with the person before sensitive actions like sending an email or making a purchase." Rush describes this in Meta Gave Every Consumer a Cloud Computer and is direct about its own gap: "Rush has nothing like Sentinel yet. Approvals that reach a person from inside the sandbox are still being built."

So the approval step has to come from the tool side, and LumisReach has two kinds of tools:

ToolApproval
start_ai_phone_call, start_ai_text_conversation, send_quick_text, send_quick_emailNone. They act and bill when called; use them for work you already decided to do
ask_lincoln (campaigns, buying numbers, bulk sends, video)Lincoln stops with awaiting_approval and the cost before anything that costs money or cannot be undone

An approval gate is only a gate if a person answers it. The agent holding the key can technically send the approval itself, so make the rule explicit in the saved agent's instructions: when Lincoln returns awaiting_approval, report the pending approvals and the cost, then end the turn. The Rush run parks with its memory intact, costs nothing while it waits, and a person resumes it with a follow-up such as "approved, go ahead." A Rush member who connects Slack gets the "needs you" message when the run stops.

For a hard stop rather than a convention, give the unattended agent a read-only key (calls:read, sms:read, contacts:read) so it can build the list and the message, and let a person launch the send from LumisReach.

Layer 3: what counts as proof the call happened

Rush's post Done Means Installed and Demonstrated lays out a ladder for code (PR opened, checks green, merged, released, installed, demonstrated) and asks every report to name the highest rung it actually reached, because "A report like that is cheap to write and expensive to fake." Calls have a ladder too, and LumisReach records each rung on the server:

RungWhere it is recorded
QueuedThe quick-action job, with one target per person
Dialedlive_call_status while it rings, from_number for the line used
Reachedend_reason of answered, voicemail or max_attempts, or a no_answer status
Goal metend_reason of goal_achieved, plus the AI summary
DemonstratedThe speaker-labelled transcript and the recording, on the contact

Ask the agent to report each person's rung with the call id, not just "done." Then check it the way Rush's post Never Let the Author Grade Its Own Work recommends for code: let a second session, which sees only the task and the call records and never the first agent's summary, confirm the outcomes. It reads the transcripts through get_call for free.

This matters because Rush's own proof only covers code today. In The Harness Is Free Now, Rush says a coding run that cannot show its pull request ends needs_review, while for other work, "checking that kind of output against the request is still ahead of us." For phone work, the LumisReach call record is the proof.

Layer 4: a record outside the agent

Rush's post When You Can't Read the Model's Thoughts, Read Its Actions sets three requirements for an action record: it "sits outside the agent," it is complete, and it is attached to one unit of work. It also says to "Meter what one finished task cost," not the price per token.

Run on Rush with LumisReach, there are two such records and neither lives inside the sandbox. Rush stores each session's trajectory, tokens and cost with the session, readable through its API. LumisReach stores each call and text, its recording, transcript, outcome and the charge for it in your workspace, and the MCP server has no tool that deletes them. Together they answer the question a manager actually asks: what did this run do, to whom, and what did it cost per lead reached?

A checklist before you let it call customers

  1. A dedicated LumisReach key for Rush, scoped down where possible, on a Rush connection with bearer auth.
  2. A saved agent whose instructions set calling hours, attempts, opt-out handling and "never answer your own approvals."
  3. Quick call and text tools only for contacts who asked to hear from you; new lists go through ask_lincoln and a person's approval.
  4. A max runs per hour on any automation, and an alert when a scheduled run produces no calls.
  5. A report that names each person's rung with a call id, checked by a second session or a person.
  6. A weekly look at cost per lead reached in LumisReach usage, beside Rush's per-session cost.

The scheduled version of this setup, with recipes, is in scheduled AI calls and texts with Rush automations.

Frequently asked questions

Is it safe to let an AI agent make phone calls for my business?

It can be, if the agent runs isolated with a key scoped to the job, a person approves anything that spends money or reaches a new list, and success is judged from the call record rather than the agent's summary. On Rush with LumisReach, the run gets its own microVM, the key is checked server side on every call, and every call's outcome, transcript and cost are stored outside the agent.

Does Rush approve an agent's actions before they happen?

Not today. Rush says it has nothing like Meta's Sentinel yet and that approvals reaching a person from inside the sandbox are still being built. With LumisReach, approvals come from Lincoln AI, which stops before anything that costs money or cannot be undone; instruct the agent to end its turn on awaiting_approval so a person answers.

Can the agent see my LumisReach API key?

Rush stores a connection token encrypted and never returns it in a response. If you want the agent never to hold the key at all, call the LumisReach REST API with a Rush header secret for api.lumisreach.com, which Rush adds outside the sandbox. Either way, use a dedicated key you can revoke.

How do I prove an AI call actually happened?

Use the LumisReach record, not the agent's report. Each call has an end reason (such as goal_achieved, answered, voicemail or max_attempts), an AI summary, a speaker-labelled transcript and the recording, all readable for free through the API or MCP tools like get_call.

Sources

  • Rush: Isolation, plainly
  • Rush: Results and proof
  • Rush blog: Every Agent Demo Hides the Same Number
  • Rush blog: Agents Are Eating the Org Chart
  • Rush blog: Your Sandbox Is Only as Isolated as Its Package Proxy
  • Rush blog: The Agent Turned Off Its Own Sandbox
  • Rush blog: Meta Gave Every Consumer a Cloud Computer
  • Rush blog: Done Means Installed and Demonstrated
  • Rush blog: Never Let the Author Grade Its Own Work
  • Rush blog: The Harness Is Free Now
  • Rush blog: When You Can't Read the Model's Thoughts, Read Its Actions
  • LumisReach API authentication and scopes
  • LumisReach Lincoln AI approvals
  • LumisReach Quick Agent Actions

Want to hear it for yourself?

Book a 30-minute call and we will set up an agent for your business, live. Or create an account and start on your own for $20.

Book a DemoGet Started

Keep reading

DevelopersHow to Give a Rush Agent a Phone Line with the LumisReach MCP Server (2026)Read article DevelopersScheduled AI Calls and Texts: Running LumisReach on Rush Automations (2026)Read article Buyer's GuideFuneral Home Answering Service: Handling First Calls, Families and After Hours With CareRead article DevelopersAI Phone Call API and MCP Server: Give Your AI Agent a Phone Line (2026)Read article

Light Up your
Conversations

Products
  • AI receptionist
  • AI outbound calling
  • Speed to lead
  • AI video marketing
  • Landing page optimization
  • Custom AI agents
Developers
  • AI phone call API
  • API reference
  • MCP server
Integrations
  • Google Calendar
  • HubSpot
  • Zoho CRM
  • Zapier
  • See all
Resources
  • Blog
  • Documentation
  • Industries
  • Pricing
  • Contact
  • Book a demo
Company
  • About
  • Partners
  • Affiliate
Legal
  • Terms of Service
  • Privacy Policy
  • Security
  • SMS Subscribe
Connect
  • X (Twitter)
  • Discord
  • Instagram
  • LinkedIn
www.lumisreach.com/llms.txt
© LumisReach · Lumis Laboratories, Inc.
2026