New product concept · Enterprise

CustomGPT Envoy

Your company's agent. It answers when tagged, and works while you're away.

Mode 1

Answer

Tag it in X, Slack, Teams, Discord or Reddit. It replies from your own sources, with citations.

Mode 2

Act

Runs on schedules and events, proposes work, and acts after a human approves.

Working name · draft for discussion · Nov 2026

Why now

Consumer agents won the attention. Enterprises can't use them.

1.67M

fact-check requests sent to Grok & Perplexity on X in 7 months

Renault, Mosleh & Rand · Feb–Sep 2025
30%

of Grok answers name a source or give a link

arXiv 2610.17068
33–40%

of Grok's World Cup fact-checks wrong or only partly right

Maldita.es · Oct 2026
2.5M+

reported US downloads of Meta Muse in its first two weeks

Sensor Tower via LA Times · unverified
No one offers this grounded, governed and auditable. That's our lane.
Mode 1 · Answer

Tag it anywhere. It answers from your sources.

D
Dana Ortiz @dana_ops · 2m
@AcmeCloud does the Pro plan include SSO? The docs contradict each other.
A
Acme Cloud Assistant Automated · run by Acme
Yes. SAML SSO is included on Pro and Enterprise. Setup takes about 10 minutes.
1SSO setup guide 2Plans & pricing
✓ Grounded in 2 sourcesApproved by Maria · Support
  1. Mention detectedX, Slack, Teams, Discord, Reddit
  2. Reads the whole threadThread goes in as untrusted context
  3. Plan & Act answersKnowledge base first, web search if allowed
  4. Policy checkAuto-post, or queue for a human
  5. Posted and loggedEvery reply in the audit log
Mode 2 · Act

It works in the background, and asks before acting.

  1. TriggerSchedule, event or a standing goal
  2. Runs on its ownPlan & Act, no chat window needed
  3. Proposes actionsDraft, message, update a system
  4. Human approvesBy policy: per action, per channel, per role
  5. Acts and remembersVia integrations and MCP, outcome saved to memory
Envoy · Approvals3
Today
⏱ SCHEDULE · MON 09:00
Weekly churn-risk digest is ready
→ Post to #cs-leads in Slack
✓ Approved · posted to #cs-leads
⚡ EVENT · ZENDESK
“Export failing” tickets up 42 in 2 hours
→ Draft a status-page FAQ and notify on-call
✓ Approved · on-call notified
🎯 GOAL · Q4 RENEWALS
Globex renews in 21 days
→ Email a renewal brief to the account owner
✓ Approved · email sent
How we build it

The engine already exists. We add the doors in and out.

New

Triggers

Mentions, events. Schedules exists

New

Channel gateway

X, Slack, Teams, Discord, Reddit adapters + thread builder

Exists

Plan & Act engine

Knowledge base, web search, Smart Tasks, Skills, Personas

New

Policy & approvals

Who approves what, confidence threshold, kill switch

New

Delivery

Post, message, email. 80 integrations + 35 MCP exists

New

Memory & audit

Cross-conversation memory, full action log

≈ 60% exists in the app today40% new

Reuse Already in customgpt-app

  • Headless Plan & Act runsAgenticSessionDispatchService::dispatchHeadless
  • Scheduled Tasks, checked every minuteapp/Modules/ScheduledTasks
  • Thread context per messagecustom_context on /tasks
  • 80 API integrations, 35 MCP connectorsBuilderIntegrations · MCPOAuth
  • Per-tool confirmation already modelledrequire_tool_confirmation
  • Synced sources: sites, Zendesk, Confluence, Drive…ProjectSource · RefreshSourcesManager

Build New for Envoy

  • Channel adapters with inbound queue + dedupeX API · Slack Events · Teams Bot · Discord · Reddit
  • Approval inbox and policy rulesper action · per channel · per role
  • Delivery service for resultstoday scheduled runs just sit in history
  • Memory across conversationsnone today
  • Event triggers and one-off taskstoday: timer only
  • Audit log export + accuracy dashboardenterprise checkbox
Pros & cons

Act is closer to done. Answer is louder.

Answer mode

Grok-style, tagged replies

  • Grounding and citations, exactly where Grok is weak
  • Reuses the knowledge bases customers already sync
  • Visible, public, easy to market
  • Same engine serves Slack and Teams internally
  • A wrong public answer carries the customer's brand
  • X rules: tagged-only, one reply, “Automated” label
  • Low reach: half of Grok replies get ≤20 views
  • Every platform API differs and keeps changing

Act mode

Muse-style, background work

  • Most parts exist: Scheduled Tasks, Smart Tasks, MCP
  • Recurring value: sticky, more usage per seat
  • Approvals + audit: what Muse can't offer enterprises
  • Fits current billing via query multipliers
  • No delivery or memory today, real build needed
  • Badly timed nudges lose trust fast
  • Touches customer systems: security review, liability
  • Plan & Act is async only, slower to respond

As a business

New enterprise product

  • New SKU on an engine we already run
  • Clear story against consumer agents
  • Land in Slack/Teams, expand to public channels
  • Each phase is sellable on its own
  • Competes for focus with the core roadmap
  • Meta and xAI ship fast and give it away
  • Needs eval and safety work before launch
  • Who owns a public mistake must be in the contract
Risks & guardrails

Safe by default, loud only when allowed.

Wrong public answer
→
Review-first by default, confidence threshold, “I don't know” + hand-off
Prompt injection from public posts
→
Thread is data, not instructions. No tool actions in Answer mode
Spam flags or platform ban
→
Reply only when tagged, rate limits, Automated label
Runaway cost
→
Per-channel budget, query multiplier, kill switch
Leaking internal data
→
Sources scoped per channel, PII redaction, roles
Over-eager proactivity
→
Digests over pings, quiet hours, act only when acceptance is likely
Enterprise must-haves
SSO / SAML exists
Teams & roles exists
Approval policies new
Audit log export new
Sources per channel new
PII redaction new
Kill switch new
Accuracy dashboard new
Roadmap

Start inside the company. Go public last.

Phase 1

Act, internal

Scheduled results delivered to Slack, Teams, email + approval inbox

Mostly reuse
Phase 2

Answer, internal

Tag the agent in Slack and Teams threads

Medium
Phase 3

Act+

Event triggers, memory, one-off tasks

Large
Phase 4

Answer, public

X, Discord, Reddit. Review-first

Large
Unit costs

Links are the expensive part on X.

$0.015 / reply

Plain-text reply via the X API (pay-per-use)

Cite sources as text by default
$0.20 / reply with a link

About 13× a plain reply

One link max, only when asked
≈ $0.10 / 20-post thread

Reading the thread at $0.005 per post

Cap thread depth per channel

Model cost: Plan & Act can take up to 15 actions a turn, so Envoy needs its own query multiplier. X rates from third-party guides, check docs.x.com before pricing.

Open questions

What we need to decide next.

1Separate product, or an Enterprise add-on?
2Which design partner first, and which channel do they live in?
3Who is liable for a wrong public answer?
4Per-customer sandboxes like Muse's VM, or stay on Smart Tasks?
TechCrunch · Meta debuts Muse TechCrunch · Muse for small businesses Wikipedia · Muse (AI agent) aicybr · Muse VM and Sentinel Indicator · Grok as a fact-checker arXiv 2610.17068 · verification requests Maldita · Grok on World Cup claims Grok in the Wild · ICWSM 2026 Asking Grok · arXiv 2605.19720 Proactive Service Agents · arXiv 2609.03727 PRISM · proactive agents X Developers · reply restriction X developer guidelines X API pricing 2026