AgentX
AgentX is an enterprise platform for building, evaluating and deploying teams of AI agents. An orchestrator delegates to specialized sub-agents, evaluation gates every release, and a single agent ships to seven channels, from Slack to voice.
What is AgentX?
AgentX Inc. publishes an enterprise platform for orchestrating, evaluating, tracing and observing AI agents. Its central argument is stated bluntly on the home page: most agent platforms stop at build, and prototypes that cannot be evaluated never reach production while agents that cannot be deployed never get past a demo. AgentX therefore sells a chain of three steps, build, evaluate and deploy, rather than an agent builder alone.
The building block is a multi-agent workforce. An orchestrator coordinates members that each carry a defined role, instructions, model, memory and tools, and the final answer is the product of their teamwork. Workflows are assembled by drag and drop, with branching logic, memory, hand-off to a human and prompt templates, while a chat-to-build mode creates and edits them in plain language. You bring your own LLM or use the platform's, paying in credits that differ by model: 10 for GPT-4o, 12 for Claude Sonnet, 8 for Gemini Pro and 3 for Claude Haiku per interaction. A knowledge layer gives each agent or team a private RAG base, hybrid search with re-ranking, knowledge graphs and high-fidelity parsing of complex PDFs, spreadsheets and scans, with per-agent access control.
Evaluation is where the platform stakes its difference. Test sets are built from unstructured data, ground truth is synthesized from documents or knowledge bases, runs are repeated to measure consistency, multi-step trajectories are scored by LLM judges, and a generated analysis says what to fix. Agent CI/CD blocks deployment when evaluations fail and promotes when they pass, while runtime monitoring adds round-the-clock evaluation, drift detection and traces of every hand-off, tool call and result. Compatibility is claimed with LangChain and LangGraph, CrewAI, the OpenAI Agents SDK, AutoGen, Google ADK, LlamaIndex, LiteLLM and any OpenTelemetry-instrumented application.
Deployment spans API and SDKs, webhooks, agents exposed as MCP servers, Slack, Microsoft Teams, WhatsApp Business, an embeddable widget, per-agent email and voice. Infrastructure comes in three shapes: managed multi-region AWS cloud, hybrid with the data plane in your own VPC, or on-premise up to air-gapped. A second business model runs alongside: a managed service in which AgentX scopes, builds, evaluates, deploys and operates a business process in 30 to 60 days.
What it does
- Design multi-agent workflows visually, with an orchestrator delegating to specialized sub-agents that each carry their own role, instructions, model, memory and tools
- Test agents against datasets before release and block promotion to production when an evaluation fails
- Deploy one and the same agent to API, Slack, Microsoft Teams, WhatsApp, an embeddable web widget, email and voice
- Plug in tools: 200+ built-in integrations, 1,000+ MCP servers and custom Python tools
- Build a private knowledge base per agent, with RAG, hybrid search with re-ranking and knowledge graphs
- Monitor agents in production: traces, tool calls, hand-offs between agents, costs and latencies
- Version an agent, roll back in one click, and place human checkpoints at any step of a workflow
When to use AgentX / When not to
A quick filter to help you decide if AgentX is the right fit.
When to use AgentX
- Teams with a mandate to put AI agents into production rather than stop at a prototype, and who must prove reliability first with test sets, repeated runs and LLM-as-judge scoring
- Solo builders and small internal teams, who can start on the permanent free plan and grow into Solo Builder without a sales call
- Agencies and service teams that deploy agents under their own brand, using white-label delivery and dedicated client workspaces from the Professional tier
- Operations leaders in companies of roughly 100 to 1,500 employees who want a whole process scoped, built and run for them: document handling, customer operations, onboarding or back office
- Organizations whose procurement runs a security review and who need RBAC, SSO, audit logs, workspace isolation and a credible path to on-premise or air-gapped deployment
When not to use AgentX
- Anyone who needs a native mobile app: there is no iOS or Android application, and mobile reach goes through WhatsApp or the embeddable web widget
- Buyers who require a certification already in hand, since SOC 2 Type II is described as an audit in progress expected in Q3 2026 and ISO 27001 only as a 2027 roadmap item
- Teams that need the full evaluation program, which is reserved for the Enterprise tier while the other paid tiers only get a demonstration mode
- Users planning to run production on the free plan, capped at one workspace, five agents, a single one-off allowance of 200 credits and one seat
- Buyers who need an interface or documentation in a language other than English, or who require a published postal address, an about page and named leadership before they can sign
How to use AgentX
A typical end-to-end flow, from setup to results.
- Create a free account on app.agentx.so, with no credit card and no onboarding call required
- Pick a starting point: fork an agent from the template catalog, or build from scratch
- Define the orchestrator and its sub-agents, giving each one a role, instructions, model, memory and tools
- Wire up the tools: enable built-in tools agent by agent, install an MCP server in one click with OAuth or an API key, or write a Python tool in the built-in editor
- Load the agent's knowledge base with documents, PDFs and spreadsheets, then set its access rights
- Build a test set from real cases, run an evaluation, read the scoring and the failures, and iterate until you clear your threshold
- Deploy: choose one or more channels, version the agent, publish it, and roll back in one click if needed
- Operate: follow traces, costs and drift, and pipe the logs into your SIEM or your OpenTelemetry backend
- Go further with the developer documentation on developers.agentx.so and docs.agentx.so, or use the Python SDK to trace agents built with LangChain, CrewAI or OpenAI Agents
- Alternative route: book a call for the managed service, starting with a scoping session with the process owner, then build and evaluate, deploy, and operate
Pros & Cons
Pros
- The whole build, evaluate and deploy chain sits in one workspace, which is the difference the publisher openly stakes its positioning on
- Evaluation is a first-class part of the product, with test sets, LLM-as-judge scoring, trajectory evaluation and CI/CD gating, rather than a bolt-on
- Observability works with third-party frameworks: you can trace a LangChain, CrewAI or OpenAI Agents agent without rewriting it, and the trace and evaluation engine can be self-hosted
- Seven deployment channels from a single agent, with no need to rebuild the agent for each channel
- The security posture is documented in unusual detail and honest about what is not certified, with three infrastructure models up to air-gapped and an explicit shared responsibility table
- The same security baseline applies from the free plan to Enterprise: RBAC, audit logs, credential vault and isolation are not behind a paywall, and audit logs export to OpenTelemetry and SIEM for everyone
- A permanent free plan with no credit card and no mandatory sales call, no per-seat billing on the first paid tier, EU data residency, a named Article 27 GDPR representative, and a managed-service route for organizations with no technical team
Cons
- No security certification to date: SOC 2 Type II is an audit in progress expected in Q3 2026, ISO 27001 only a 2027 roadmap item, and the announced bug bounty is not open yet
- The privacy policy and the security page contradict each other on training: one allows training on your content with an opt-out, the other rules it out without reservation
- The company is opaque: no postal address, no about page and no named executive anywhere on the site
- No subprocessor list is published, even though the privacy policy states one is available on the site
- Credits expire at the end of each billing cycle with no rollover, and both credits and subscription fees are non-refundable
- Pricing contradicts itself between two blocks of the same page, with Professional at 199 USD per month in the plan grid and 249 USD per month in the cost examples, and the pricing FAQ answers are not served as static HTML
- No support email is published, so support goes through a form; there is no mobile app; the full evaluation program is Enterprise-only; and the site and documentation exist in English only
Pricing & Plans
AgentX offers a permanent free plan at 0 USD, covering one workspace, up to five agents, a one-off allowance of 200 credits and a single seat, with no credit card required; there is no time-limited free trial. The lowest paid entry point is Solo Builder at 49.00 USD per month, or 490 USD per year. Above it, Professional is listed at 199 USD per month (1,490 USD per year) and Business at 299 USD per month (2,990 USD per year), while Enterprise is quoted per process. Usage is metered in credits consumed per interaction at a rate that depends on the model, namely 10 credits for GPT-4o, 12 for Claude Sonnet, 8 for Gemini Pro and 3 for Claude Haiku; additional credits cost 10 USD per 1,000 on every paid tier, and extra seats 10 USD per month on Professional and Business. The publisher's own cost examples cite roughly 79 USD per month for personal automation, 129 USD for an internal team and 349 USD for client work. One inconsistency should be noted: the same pricing page shows Professional at 199 USD per month in its plan grid and at 249 USD per month in its cost examples. Credits are non-transferable, non-refundable and expire at the end of each billing cycle, and subscription fees are non-refundable except where the law requires otherwise.
- one workspace
- up to 5 agents
- a one-off 200 credits
- 1 seat
- multi-agent workflows and API access
- unlimited workspaces
- up to 25 agents
- 5
- 000 credits per month
- extra credits at 10 USD per 1
- 000
- 1 seat
- production deployment
- unlimited workspaces
- up to 25 agents
- 10
- 000 credits per month
- 2 seats with extra seats at 10 USD per month
- white-label deployment
- client workspaces and the evaluation demonstration mode
- unlimited workspaces
- unlimited agents
- 20
- 000 credits per month
- 2 seats with extra seats at 10 USD per month
- white-label
- client workspaces
- priority support and a priority SLA
- unlimited workspaces
- agents and seats
- custom credit volume and pricing
- white-label
- the full evaluation program (agent risk mapping
- custom evaluation sets from real cases
- multi-run reliability testing
- benchmarks against accuracy targets
Data, GDPR & hosting
A consolidated view of how AgentX handles your data.
GDPR overview
GDPR is addressed concretely, but alignment is not certification. The privacy policy lists legal bases for EEA, UK and Swiss users (contract, legitimate interests, consent, legal obligations) and the rights of access, erasure, rectification, portability, restriction, withdrawal of consent, objection, complaint to a local authority and protection against automated decisions. Requests go to contact@agentx.so and appeals to legal@agentx.so, after identity verification; authorized agents are accepted on signed written proof. An Article 27 EEA representative is named: dr Poket Poland Sp. z o.o., Leborska 3B, Gdansk, Pomeranian District, 80-386. Transfers outside the EEA, UK and Switzerland rely on standard contractual clauses, and a GDPR-aligned DPA template is available on request. Confirmed incidents affecting customer data are notified within 72 hours. DORA, the EU AI Act, SR 11-7, MAS, HKMA and SOX are listed as alignment frameworks only.
Who owns the data?
The terms leave your content with you, but grant AgentX Inc. a broad license over it: the company may copy, modify, create derivative works from, publish, distribute and incorporate that content into its own products. Feedback you send is claimed even more broadly, for commercial or non-commercial purposes. External API keys you supply are collected, stored and treated as sensitive, and the privacy policy states you may view, update or delete them at any time from your account. Portability is offered: on request, the company provides a machine-readable JSON, CSV or ZIP export of your account information, chat history, files and billing metadata within 30 days.
Reuse rights
Two AgentX documents disagree about training, and both are live. The privacy policy, last updated 15 November 2024, carries a box titled "No Retention for Generalized AI/ML Model Training" stating that no user data is retained or used to develop, improve or train generalized AI or machine learning models. Yet its legitimate-interests section says the company may use your Content to improve the service, "for example, to train or fine-tune the models that power AgentX", with an opt-out on request to contact@agentx.so. The security page, updated 13 May 2026, is unqualified: "No customer data used to train AgentX models - full stop. Documented and enforceable." It also states that under the default configuration customer conversations are never sent to LLM providers' training pipelines. Data handled through the Google Workspace API is limited by the Google User Data Policy and never used to train generalized models, and the company says data passing through its OpenAI integration is not used to train or improve OpenAI's models. Declared purposes are service delivery and administration, improvement and research, communication, development of new services, fraud prevention and security, business transfers and legal obligations; aggregated or de-identified data is used for analysis and research with no attempt at re-identification. Enterprise customers may rely on their own commercial agreements with Anthropic, OpenAI or Google for stronger terms. Ask the publisher to reconcile the two texts rather than assuming either one governs.
Data retention & training
Hosting summary
The privacy policy states that personal information is processed and stored in the company's own facilities and servers in the United States, and may be disclosed to service providers and affiliates in other jurisdictions. The security page adds options. On the managed cloud tier, hosting runs on multi-region AWS with private subnets, security groups, no public access to internal components and AWS Shield DDoS protection; EU residency is available, with data staying and being processed in European regions, and US residency is offered as well. In hybrid deployments the data plane stays inside the customer's own VPC on AWS, Azure or GCP, while AgentX manages the control plane over signed and encrypted channels. On-premise runs entirely inside the customer's perimeter, in a datacenter or air-gapped, with no external dependency for basic operation. In both cases the customer controls every data location. Data flows to LLM providers are documented provider by provider, and customers can pin specific regions such as Anthropic or OpenAI EU endpoints. Encryption is TLS 1.3 in transit and AES-256 at rest under AWS KMS, with customer-managed KMS on Enterprise. Transfers outside the EEA rely on standard contractual clauses.
Things to keep in mind
Risks and trade-offs to weigh before adopting AgentX.
- No certification is in hand: SOC 2 Type II is an audit in progress expected in Q3 2026 and ISO 27001 a 2027 roadmap item. The publisher states this itself, and alignment with a framework is not certification against it
- The privacy policy (updated 15 November 2024) and the security page (updated 13 May 2026) contradict each other on training: one allows use of your content with an opt-out, the other rules it out without reservation, and both are online at once
- The privacy policy states that a subprocessor list is available on the site, yet no such page exists: /subprocessors returns a 404 and it is absent from the sitemap
- No postal address and no named executive appear anywhere on the site, which makes the counterparty hard to assess before a purchase or a data transfer
- The terms grant the publisher a broad license over your content and over any feedback you send, for commercial or non-commercial purposes
- Disputes go to mandatory individual arbitration in San Francisco with a waiver of class actions; credits expire each cycle without rollover, subscription fees are non-refundable, and the same page shows the Professional plan at two different prices, 199 and 249 USD per month
- Commercial figures on the site cannot be verified independently, including 10M+ interactions per month, a 99.97% SLA and 60-80% of inbound resolved automatically. Treat an evaluation score as a measurement against your own test set rather than proof of judgment, and keep a human checkpoint wherever the decision actually matters, or the fluency of a passing agent will quietly replace your own review
Setup & Integrations
Technical difficulty
Low to start, higher to industrialize. The publisher promises a first agent in 10 minutes, with no credit card and no onboarding call: workflows are built by drag and drop or in natural language, templates can be forked, and MCP integrations install in one click with OAuth or API-key authentication handled. Complexity arrives later: custom Python tools, the Python SDK, tracing from a third-party framework, CI/CD gating. Hybrid and on-premise deployment assumes an infrastructure team for VPC, IAM, KMS and patching. Organizations with no technical team can take the managed-service route: 30 to 60 days to production.
Deployment
Integrations
Behind AgentX
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Do I need a credit card to get started?
What does the first paid tier cost?
How are the models billed?
Do unused credits roll over to the next month?
Which channels can an agent be deployed to?
Can AgentX run on our own infrastructure?
Is the platform SOC 2 certified?
Will my data be used to train models?
Where is the data hosted?
How long is my data kept after I delete my account?
Should you pick AgentX?
AgentX earns attention for the chain, not for the builder. Plenty of platforms let you assemble a multi-agent workflow; far fewer make evaluation a gate, with test sets drawn from real cases, repeated runs, LLM judges and a CI/CD rule that refuses to promote an agent that fails. Add seven deployment channels from a single agent, tracing that works on LangChain or CrewAI agents built elsewhere, and three infrastructure models up to air-gapped, and the product reads as one designed by people who watched prototypes die between demo and production.
Three doors lead in, and they suit very different buyers: self-service for solo builders on a permanent free plan, white-label with dedicated client workspaces for agencies from the Professional tier, and a managed service in which AgentX scopes, builds and operates an entire process in 30 to 60 days for operations teams with no engineers to spare.
The reservations are real. Nothing is certified yet: SOC 2 Type II is an audit in progress for Q3 2026 and ISO 27001 a 2027 target. The publisher says so itself, which is to its credit, but alignment is not certification. The privacy policy and the security page contradict each other on whether your content may train the models, and both texts are online at once. The company publishes no address, no about page and no named leadership. Credits expire every cycle without rollover and nothing is refundable. The pricing page even disagrees with itself, showing Professional at 199 USD in one block and 249 USD in another.
Worth a serious trial if you need agents in production and have to prove reliability before release. Before signing, ask the publisher to reconcile its training language, to publish the subprocessor list its own policy promises, and to confirm which Professional price applies.
- Choosing a selection results in a full page refresh.
- Opens in a new window.