Hyperagent in Slack: Build AI Teammates Your Team Will Use

Direct Answer

Hyperagent's strongest idea is not that an AI becomes an employee. It is that a team can reach a named, persistent, tool-using agent from Slack without learning an agent console first. In Andrew Warner and Alex McDonnell's demonstration, a sales agent called Signal researches a company, enriches contacts with Apollo, creates a source-backed prospect page, drafts Gmail outreach, updates Airtable, and watches for replies on a recurring loop.

That is a useful team interface. It is not automatic proof that the workflow is safe, economical, or ready to send messages unattended. A production version needs four layers the metaphor can hide: a narrow job, least-privilege tools, explicit approval gates, and an outcome log that determines what the agent is allowed to remember as a successful pattern.

Watch the Hyperagent Team Demo

Credit to Andrew Warner, Alex McDonnell, and The Next New Thing for the walkthrough. The supplied Hyperagent campaign link advertises $500 in bonus credits on an eligible paid plan. Detailed public eligibility was not visible during this review, so verify the amount, plan requirement, expiry, renewal, and overage terms at checkout. The episode's companion links are collected on Andrew's resources page.

Link Map

The Architecture in One View

Hyperagent documents three core objects: a thread records one work session, an agent carries identity, instructions, tools, and knowledge, and learnings hold reusable skills, memories, and rubrics outside the active conversation. The Slack bot is another invocation surface for that configured agent, not a separate intelligence.

What the Signal Sales Sprint Actually Does

Alex starts a roughly 20-minute outbound sprint from Slack using a public company as the target. Signal researches the business and possible buying signals, finds responsible contacts, packages its findings into a command center, creates a personalized page, and writes concise Gmail drafts. Airtable acts as the lightweight CRM; Calendar is available for meetings.

The best part of the demo is the artifact, not the volume. The page links to public sources and lets the seller inspect why the system believes a project is relevant. That turns a vague AI recommendation into reviewable evidence. The weak point is the jump from public evidence to contact permission. A company announcement can support relevance; it does not automatically make unsolicited outreach appropriate.

Hyperagent's current terms explicitly prohibit using the service to send unsolicited communications, promotions, advertisements, spam, or unsolicited, abusive, or deceptive messages. A compliant pilot should therefore use warm inbound leads, existing relationships, event opt-ins, referrals, or another approved first-party list. For a fuller outreach control model, see the signal-based marketing-agent guide.

Why Slack Is More Than a Convenience

Agent projects often fail at adoption because one technical person owns the terminal, credentials, and undocumented operating habits. Putting the invocation and review loop in Slack changes who can participate. A teammate can mention the SEO agent, inspect the result, ask a follow-up, and see the decision history without learning a new builder interface.

The video shows three workspace roles inside Hyperagent: owner, editor, and member. It also notes that people already inside an allowed Slack channel can interact with its bot even if they do not work in the Hyperagent web app. That makes channel membership part of the effective access model. Put finance, customer support, sales, and public-content agents in separate channels; do not treat one broad company channel as a universal permission group.

A Named Agent Is Not a Permission Boundary

A separate name and avatar are excellent cognitive design. They tell a person which agent handles sales versus SEO. They do not by themselves stop the SEO agent from reading Gmail or the sales agent from touching a calendar. The real boundary is the combination of enabled tools, connected account scopes, knowledge access, Slack channels, model budget, allowed delegate agents, and unattended autonomy settings.

Hyperagent's current configuration reference exposes budgets, trigger history, tool toggles, integration access, memory scope, and knowledge presets. Its July autonomy update also separates what an agent may do in a supervised thread from what it may do on a schedule or Live Mode, and lets an agent delegate only to allowed agents. Those are the controls to review before admiring the avatar.

A Practical Approval Matrix

Personalized Pages Need a Restraint Layer

Signal's personalized pages combine company research, project announcements, product fit, animation, and an inline video. This can be genuinely useful when the page answers a prospect's real question with traceable sources. It becomes uncomfortable when the personalization feels like surveillance, overstates the evidence, or publishes personal details merely because they were easy to find.

• Use company-level, role-relevant public information rather than personal-life details.
• Show source links and the date each claim was checked.
• Label estimates and inferred fit instead of presenting them as known facts.
• Keep the first version private and require a reviewer to approve every claim.
• Give the page an expiry date and an unpublish path.
• Do not upload documents or personal data unless the organization has the right to process them.

Live Mode Is a Stateful Loop, Not Magic Autonomy

In the recording, Signal's Live Mode repeatedly checks inboxes, prospecting work, nurture tasks, and learning instructions in the same thread. Current Hyperagent documentation distinguishes this stateful Live Mode from scheduled runs that create a new thread. The current cadence options include every 2, 4, 8, or 12 hours, daily, or a custom interval, with an estimated daily cost before saving.

That difference matters operationally. A persistent thread can carry state, but it can also accumulate a mistaken assumption. A fresh scheduled run is easier to isolate, but it needs an explicit state store. For either mode, define a stop condition, maximum records per run, maximum spend, quiet hours, duplicate protection, and an alert path. Hyperagent also supports alert-only schedules, which is a sensible first recurring mode: say nothing unless a defined exception appears.

Self-Improvement Needs Reviewed Memory

The demo shows Signal sharing hypotheses with a marketing agent and updating skills or memories based on replies. This is where multi-agent systems can compound value, but also where one bad inference can spread. A reply is not automatically proof that a message, channel, or claim caused the result.

Hyperagent now links memories back to their source thread and offers deduplication and bulk review. Use that provenance. Promote a learning only when it includes the source run, sample size, outcome definition, reviewer, and expiry date. Keep a separate approved playbook from experimental hypotheses, and allow rollback when a memory degrades later work.

Security Starts With the Connector Map

A sales agent can touch email, contacts, calendar, CRM records, web sessions, and research services. That is a large credential surface. Hyperagent's privacy policy says it uses encryption in transit and at rest, but correctly makes no absolute security guarantee. Its terms also identify multiple AI and infrastructure subprocessors and state that third-party AI providers may retain inputs and outputs for up to 30 days for safety and compliance moderation.

The connector history is also worth knowing. Hyperagent reported that a May 2026 incident at integration provider Composio potentially exposed a small number of customer GitHub OAuth tokens and may have affected other stored tokens. Hyperagent disabled Composio-powered integrations, verified revocations where possible, notified affected customers, and launched first-party Google, GitHub, and Notion integrations alongside its existing native Airtable and Slack connections. That is a responsible response, and also a reminder that integration vendors sit inside the security boundary.

1. Use a test workspace, test Airtable base, and narrow agent mailbox first.
2. Inspect every OAuth scope and prefer current first-party integrations.
3. Do not share passwords; use the supported authorization flow.
4. Review connected-app and provider audit logs after the pilot.
5. Revoke unused access at the provider, not only inside Hyperagent.
6. Keep regulated, highly sensitive, and prohibited data out of the workspace.
7. Assign one human owner for credentials, incidents, and agent shutdown.

The ChatGPT Connection Does Not Make the Workflow Free

Near the end of the video, Alex demonstrates Settings > AI Providers and connects a ChatGPT subscription so that account is used before Hyperagent credits. He also says the workflow still incurs smaller fees for services such as Exa search and Browserbase. This is a creator-demonstrated feature; public product documentation did not provide a complete entitlement contract during this review.

Budget for the whole system: the eligible Hyperagent plan, model or subscription limits, search, browser sessions, media generation, Apollo, Airtable, Google Workspace, Slack, and human review. Hyperagent's current model picker reports token use, caching, and running thread cost, while its terms allow plans, allotments, fees, and overages to change. Compare cost per approved meeting or accepted deliverable, not cost per generated page.

Hyperagent Versus Hermes Agent

The video fairly identifies the adoption advantage of a hosted Slack experience, but it overstates the idea that local agents require one computer per personality. Multiple agents can share local or remote infrastructure. The real tradeoff is managed team access versus self-managed control, not one agent versus many laptops.

Copy-Ready Signal Contract

You are Signal, a research and sales-support agent.

Objective:
Help the team identify and prepare evidence-backed follow-up for
approved, permissioned prospects. Optimize for qualified conversations,
not message volume.

Allowed inputs:
- Approved CRM view: [VIEW NAME]
- Approved inbox label: [LABEL]
- Public company sources
- Approved product, proof, and pricing documents

Allowed actions without approval:
1. Research public company facts and save source URLs with dates.
2. Draft a short fit memo and private prospect page.
3. Draft an email in Gmail.
4. Suggest CRM field updates and meeting times.
5. Flag missing evidence, ambiguity, objections, or suppression requests.

Actions requiring human approval:
- Send any message
- Publish or share a prospect page
- Enrich a new person
- Change CRM stage
- Book or move a meeting
- State pricing, guarantees, legal terms, or customer commitments

Never:
- Contact a person outside the approved list
- infer sensitive traits or use personal-life details
- hide AI involvement or invent evidence
- contact anyone marked unsubscribe, do-not-contact, complained, or bounced
- promote a learning without source run, sample size, reviewer, and expiry

Run limits:
- Maximum [N] prospects per run
- Maximum [COST] per run
- Stop after [TIME]
- Stop immediately on authorization failure, duplicate detection,
  ambiguous consent, policy conflict, or missing source evidence

Return:
- Prospect and reason for inclusion
- Source table with URLs and checked dates
- Draft artifact link
- Gmail draft link
- Proposed CRM changes
- Risks and required approvals
- Run cost and tool usage

A Seven-Day Team Rollout

1. Day 1: choose one low-risk result, one owner, one reviewer, and ten realistic test cases.
2. Day 2: create a named agent with one job, one Slack test channel, and mention-only invocation.
3. Day 3: connect only public research and a test Airtable base; no personal mailbox.
4. Day 4: add a narrow Gmail label or agent mailbox in draft-only mode. Run the ten cases.
5. Day 5: grade evidence accuracy, duplicate handling, tone, correction minutes, and cost per accepted draft.
6. Day 6: add a supervised Live Mode checklist or an alert-only schedule, not both.
7. Day 7: review permissions, memory suggestions, audit logs, provider access, spend, incidents, and team adoption. Expand only one boundary.

Success is not "the agent ran." Success means the team used it, reviewers accepted a high share of its work, permissions stayed within policy, and the total correction time plus tool cost beat the previous process.

Video Chapters

Bottom Line

Andrew and Alex show a genuinely useful adoption pattern: people and agents working in the same Slack channels, with visible identities, persistent skills, connected tools, and reviewable artifacts. That can move agent work from one enthusiast's private setup into a repeatable team process.

The winning unit is not an AI employee. It is a bounded team workflow with a clear owner, narrow credentials, source-backed output, human authority over external actions, and memory that can be inspected and rolled back. Start with one agent people actually mention, one draft they actually approve, and one metric that proves the workflow deserves another permission.

Sources and Useful Links

• Andrew Warner and Alex McDonnell: An agent you and your team will actually use
• The Next New Thing resource library
• Hyperagent promotional link supplied with the episode
• Hyperagent: threads, agents, and learnings
• Hyperagent: agent configuration reference
• Hyperagent: agents and autonomy update
• Hyperagent: memory and learning update
• Hyperagent: Composio security-incident response
• Hyperagent terms of service
• Hyperagent privacy policy
• Apollo, Airtable, Gmail, and Google Calendar
• JQ AI SYSTEMS: How to Build a No-Code AI Agent Team With Hyperagent
• JQ AI SYSTEMS: Two AI Marketing Agents for Outbound and Organic Growth