Quick answer#
AgentTeams is worth a limited pilot if you want hosted AI coworkers across your existing business tools, but its $29 starting price does not buy unlimited work. The current offer includes a free plan with 1,000 monthly credits, Starter from $29 for 3,000 credits and Business from $199 for 25,000 credits; Enterprise is custom-priced.[14] You pay for model work rather than the number of agents, and concurrency, account access and governance differ by plan.[14]
Its most useful documented features are cross-tool tasks, event-triggered work, per-action approval proposals and an activity record—not the idea that software literally replaces an employee.[18][27][29] The important limits are a variable credit bill, nonrefundable fees, a young contracting company and security controls that have not all reached the certified or available stage.[15][17][28]
- Best for: a small operations or support team willing to pilot one bounded workflow, review outbound actions and accept a hosted vendor.
- Avoid if: you require self-hosting, a completed SOC 2 report, guaranteed regional residency or an independently demonstrated uptime/performance record before connecting business data.
- Buying decision: start with synthetic data on Free. Upgrade only after measuring useful completed work, approval time and credits—not after counting how many agents you can create.
Verified October 11, 2026. We reviewed current pricing, terms, privacy/security pages, integration documentation and public third-party references. We did not register, connect customer accounts, execute product tasks, test refunds or audit the runtime. Performance and security statements below are attributed documentation, not our test results. This site promotes Hermes Agent and FlyHermes; they are relevant alternatives for different needs, not automatic winners.
Identity: which AgentTeams is this?#
This review covers agentteams.com, the hosted AI Employee Platform. The legal notice identifies AgentTeams, Inc. as a Delaware corporation incorporated September 24, 2026, with a New York contact address.[17] Those are the company's published particulars; we did not obtain an independent registry certificate.
Name matching matters. Search results also return “agent teams” as a general architecture and similarly named products on other domains. Their open-source licenses, free/pro quotas or reviews should not be imported into this company's dossier.
A third-party Recent Funding profile links to this exact domain, but it is a company/funding directory, not a product test.[33] We did not establish a substantial independent customer-review record. That gap is particularly important for a service asking for access to email, support tickets and internal documents. It does not establish fraud; it makes a low-risk pilot and a clear contract more valuable.
What the platform does#
AgentTeams describes a Coordinator and specialist agents with roles, directives, knowledge and connected accounts. Its onboarding guide says users sign in with a work Google account, define the first job with the Coordinator, connect tools and run a first task.[29] Agents can be grouped into teams, with the Coordinator handling unassigned work.[29]
The documented task types are Once, Repeat and Reaction: one-time jobs, scheduled work and responses to events in connected services.[29][31] The integration catalog includes communication, workspace, ticketing, development and business tools; it describes credentials and permissions scoped per agent rather than one universal service identity.[18]
Examples of sensible first jobs are a draft-only support response, an internal status summary or a report assembled from a restricted document set. These are our evaluation suggestions, not tasks we ran. Do not start by granting an agent broad authority over refunds, ad spend or production repositories.
The vendor's “five minutes” hiring claim is not a deployment benchmark, and its help guide separately describes getting started in under 15 minutes.[6][29] Neither figure includes an independently measured security review, permission cleanup, knowledge preparation or acceptance test.
Pricing: plans, credits and concurrency#
The live USD monthly pricing was:
- Free: $0, 1,000 credits per month, two user accounts and one concurrent task. The page includes integrations, custom connectors, webhooks and API access.[14]
- Starter: $29 for 3,000 credits; larger sizes shown at $49 for 5,000, $89 for 10,000 and $169 for 20,000. Unlimited agents and user accounts, but only two tasks at once.[14]
- Business: from $199 for 25,000 credits, with larger sizes up to 100,000. Four concurrent tasks, hands-on onboarding, per-agent spending controls and priority support.[14]
- Enterprise: custom pricing and capacity, with features including SSO, audit export, team budgets and an SLA. Connecting your own Claude or ChatGPT subscription is an Enterprise arrangement, not a Starter entitlement.[14]
The page advertises 20% savings for yearly billing and supports USD, EUR, GBP and CAD. Taxes are calculated at checkout.[14] We did not verify a personalized annual invoice or local-currency checkout. Monthly plan figures are not the total price of an annual commitment.
We encountered two public pricing layouts with the same main entry prices. The privacy policy explicitly says the company is testing two versions and remembers the version in a cookie.[16] Save the plan summary and terms you actually receive rather than relying on a search snippet or a screenshot from another visitor.
Credits: a useful cost example, not a benchmark#
AgentTeams says credits depend on how much a task reads, thinks and writes. Its examples average about 20 credits for a quick reply, 70 for an everyday task and 280 for longer research on Standard. Light uses roughly half the Standard rate, and Expert roughly double.[14]
Using those published estimates, 3,000 credits cover 42 complete 70-credit tasks, or 21 complete 140-credit tasks, before any other usage. Those are illustrative divisions, not measured throughput. Actual jobs can use more or less, and the Coordinator or other work may also consume the plan's allowance.
The site's calculator illustrates two agents doing 68 tasks each at about 70 credits: 9,520 credits a month, pointing toward the 10,000-credit Starter size at $89 rather than the $29 entry size.[14] That is a vendor workload assumption. Your pilot should record the actual meter before you choose a plan.
A sensible measure is cost per accepted result. Record total credits, failed or repeated attempts, human corrections and time spent approving. An agent that produces inexpensive drafts you must rewrite may be worse value than one doing fewer useful jobs. For another hosted usage model, the Zapier Agents review explains why activity counts should not be confused with completed business outcomes.
Renewal, automatic upgrades and refunds#
The October 2026 terms specify advance monthly or yearly billing, automatic renewal and generally nonrefundable fees except where law requires otherwise. They explicitly include unused credits and partially used periods in the no-refund rule.[15]
An admin can enable automatic upgrades. When enabled, exhausting credits moves the organization to a larger size and charges the difference without another approval each time.[15] The pricing page describes an upper size you can allow; confirm that ceiling in your actual Billing settings before enabling it.[14] “No surprise bills” should not be read as “no additional charges under any circumstance.”
Without credits, agents pause. The terms say a task already in progress finishes and its excess usage is deducted from the next credits received.[15] Ask how that state appears to external users: the pricing material describes quiet pauses on outside channels, so do not assume every customer gets a visible fallback message.[14]
Other details worth saving:
- Starter and Business carry unused monthly credits into the next month, capped at the plan's monthly allowance; carried credits expire at that month's end.[15]
- A downgrade or plan cancellation takes effect at the end of the current period. Closing the account ends the paid plan immediately, without a refund for the remaining period.[15]
- The larger size remains the renewal size after an upgrade. The terms describe the charge from the start of the current billing month, so do not assume a last-day upgrade costs only one day's difference.[15]
- After account termination, data is available for export for a “reasonable period,” not a specific number of days in the public terms.[15]
A monthly pilot limits commitment. It does not create a money-back guarantee.
Approval gates: the strongest feature to validate#
The help documentation describes per-action proposals with the intended payload and reasoning. A human can approve, edit or dismiss; pending proposals appear in chat and Tasks to review.[27] Customer-facing replies, destructive operations and financial actions are listed as gated by default. Internal reads, drafts and replies to teammates can run without gates unless you tighten the policy.[27]
That is more concrete than a generic “human in the loop” promise. It still needs testing. The documentation says an agent creates a proposal when it decides approval is needed; our review did not establish how every connector technically enforces that boundary.[27] Ask whether policy is enforced outside the model, which actions are covered and whether retries preserve the original authorization.
The activity log is described as append-only, with payload, time and approver attribution retained during the subscription.[27] An in-product log is not the same thing as export rights: audit export is listed in Enterprise pricing.[14] If you need to retain incident evidence independently, get the export format, availability and retention commitment before connecting sensitive systems.
For a configurable workflow alternative, the n8n AI Agent Builder review covers a different approach: designing the steps and permission boundaries yourself rather than starting with hosted employee roles.
Privacy and security: planned is not certified#
AgentTeams states that it uses AES-256 at rest, TLS in transit, encrypted integration credentials, Google Cloud infrastructure and per-agent permissions. Its security page says it is working toward SOC 2 Type I, that data-residency options are coming soon, and that regular third-party assessments are planned.[28] Do not convert those roadmap statements into a completed audit or an available regional-hosting guarantee.
The privacy policy covers account information, agent messages and knowledge, OAuth tokens, usage data, logs and billing records. It says customer data is not used to train models, and specifies limited use of Google Workspace data.[16] That is not zero processing: model infrastructure and connected third parties still need data to perform the work.
The policy says deletion or anonymization occurs within 30 days after account deletion, subject to legal retention, and that conversation/knowledge data is deleted when the relevant agent or organization is removed.[16] Obtain the practical backup-deletion and export procedure if those obligations matter to your organization.
Before a sensitive deployment, request a DPA, subprocessor/model-provider details, incident-notification commitments, retention rules and proof of the controls you require. A privacy page is a useful starting point, not a completed security assessment. The Kore.ai Artemis review is a more relevant comparison for buyers evaluating enterprise control architecture than a simple lowest-price ranking.
Hidden costs and remaining limitations#
The subscription is only one part of operating the workflow:
- Connected tools may require their own licenses or suitable service accounts.
- Administrators must approve OAuth scopes and maintain access as staff and responsibilities change.
- People still prepare knowledge, review outputs and investigate failed or duplicate actions.
- Model tier, document volume, retries and task frequency can change credit demand.
- Enterprise requirements can move you beyond self-serve pricing before usage becomes large.
- Export, retention and switching costs matter when an agent accumulates business context.
These are cost categories to investigate, not additional AgentTeams fees we verified. No open-source or self-hosted edition of this exact service was established in the reviewed offer. The public terms describe the service's code as owned by AgentTeams.[15]
“Unlimited agents” does not mean unlimited parallel capacity or no supervision. Likewise, a role description is not a trained human's professional responsibility. You remain responsible for the accounts you connect and the decisions made using the output.
Independent reviews, complaints and advertising#
Fresh exact-domain searches and Reddit-focused searches did not recover a substantial independent product-test or complaint corpus. The domain-matched company directory is background evidence, not evidence that tasks succeed or a security claim has been audited.[33] Unrelated “agent teams” projects were excluded.
We did not verify the landing page's hiring-time or productivity statistics. They should not become your financial forecast.[6] A safe review can recognize specific documentation and a testable free tier without pretending those are proven customer outcomes.
Ad-oriented searches and the local Meta/YouTube archive check did not yield a confirmed AgentTeams campaign ID, dates or spend. The privacy policy says the public website uses Google Ads conversion measurement after consent.[16] That alone does not prove that a particular paid campaign is currently running. Product pages about managing Google Ads are also not advertising-library evidence.
Alternatives: choose control, service or team workflow#
AgentTeams: the attraction is a hosted team interface, business-tool connections and reviewable actions.[18][27][29] It may suit a team that wants to configure responsibilities rather than maintain a runtime. Its case is strongest after a pilot proves those responsibilities are actually enforced.
A defined deployment service: if nobody on your team can scope the process or own the rollout, a software subscription may leave the hardest work undone. The Agentra review examines a service-led alternative and the need to specify implementation and ongoing responsibilities in the agreement.
Hermes Agent: official documentation describes an open-source agent with tools, memory and reusable skills; its self-hosted Web Dashboard includes real browser chat as well as administration.[7][8] The distinction is not “AgentTeams has a UI and Hermes does not.” It is who runs the system, connects the tools and validates the workflow. Use the Hermes installation guide if you want that control and can maintain it. This is not evidence of parity with AgentTeams' team-wide approval and audit features.
FlyHermes: the managed service offers browser/mobile access to Hermes without the same server-maintenance burden.[23] Compare the self-hosted versus hosted guide when uptime and maintenance are the actual problem. Managed access is not a completed compliance review or a replacement for validating business actions.
If you want a broader examination of Hermes' own benefits and limits before choosing, read the Hermes Agent review. An alternative should win on your requirements, not because this website sells the next step.
A proposed pilot before paying annually#
This is an evaluation plan, not a test we performed:
- Choose one job: for example, draft a support reply from a small approved knowledge set. Define what counts as correct and when the agent must decline.
- Use synthetic records and restricted accounts. Do not upload customer data merely to see whether onboarding works.
- Test missing context, contradictory instructions, revoked access, duplicate events and an attempted financial action.
- Confirm that an outbound action cannot occur before approval, that editing changes the actual payload and that dismissal prevents execution.
- Inspect the activity record. Can a reviewer reconstruct the request, tool action and approving human without asking the agent to explain itself?
- Record credits and human review time for both successes and failures. Check what pauses at the spending boundary.
- Export the records you need, disconnect a tool and verify access is revoked. Ask what remains after cancellation versus account deletion.
Our verdict: the documentation is specific enough to justify a bounded Free-plan evaluation, not broad production trust by default. Buy a larger plan when it delivers useful accepted work within a known budget. Choose another route if security evidence, residency, portability or operational ownership does not fit your requirements.
Sources#
[6] https://agentteams.com/ai-employee-platform — AI Employee Platform, Hire AI Workers | AgentTeams [7] https://hermes-agent.nousresearch.com/docs — Hermes Agent Documentation | Hermes Agent [8] https://hermes-agent.nousresearch.com/docs/user-guide/features/web-dashboard — Hermes Web Dashboard | Hermes Agent [14] https://agentteams.com/pricing — Pricing, AgentTeams [15] https://agentteams.com/terms — Terms of Service | AgentTeams [16] https://agentteams.com/privacy — Privacy Policy | AgentTeams [17] https://agentteams.com/legal-notice — Legal Notice | AgentTeams [18] https://agentteams.com/integrations — Integrations, AgentTeams [23] https://flyhermes.ai — FlyHermes.ai — Your Hermes Assistant [27] https://agentteams.com/help/approval-gates — Approval gates | AgentTeams Help [28] https://agentteams.com/security — Security | AgentTeams [29] https://agentteams.com/help/getting-started — Getting started | AgentTeams Help [31] https://agentteams.com/help/tasks — Tasks & schedules | AgentTeams Help [33] https://recentfunding.com/companies/agentteams — AgentTeams is an AI workforce platform | Recent Funding