IBM watsonx Orchestrate
Platform for building, running and governing AI agents, with scheduled and multi-channel triggers, per-agent model choice, imported knowledge bases, a pre-deployment evaluation framework and request traces.
IBM watsonx Orchestrate is a platform for building, running and governing AI agents. Agents can be built without code or with the Agent Development Kit's Python SDK and command line, and agents built elsewhere can be imported, including from Langflow. Agentic workflows assemble them from documented node types: handing a step to another agent, branches, parallel paths, loops, decisions, timers, and user activity steps that pause for a person to approve or supply information before the flow resumes.
Agents ground their answers in knowledge bases the customer imports, backed either by a built-in vector store or by the customer's own Milvus or Elasticsearch instance, and they can persist what they learn through a memory API that writes, searches, retrieves and deletes. Models are imported from multiple providers and assigned per agent, with policies for load balancing, fallback and retries.
Before release, the evaluation framework replays simulated conversations against reference data and reports where the agent left the expected path; in production, traces show the full route each request took, readable from the command line, the SDK, the interface or the API, and exportable over OpenTelemetry. Agents reach users through web chat, voice, Slack, Microsoft Teams, SMS, WhatsApp, Facebook Messenger and a Genesys connector, or run on a schedule. It is sold in three plans with published entry pricing and runs managed on IBM Cloud or AWS, or in the customer's own environment.
Vendor details
Canonical URL
https://www.ibm.com/products/watsonx-orchestrate
Category
Enterprise operations agent
Subcategory
Agent orchestration platform
Company status
first party product
Use cases & customers
Primary use cases
Target customers
In practice
Your teams have built agents in Langflow, LangGraph and in-house tools, and nobody can see them all. Orchestrate imports them, runs them behind one control plane, and shows the path each request took through them.
The same work comes round every morning. You mark an agent or an agentic workflow schedulable, then set the timing in plain language in the chat, and it runs without anyone starting it.
A step in the process needs a person's sign-off. A user activity node holds the workflow, asks a named approver for the decision or the missing fields, and resumes with a record of what was approved.
Agentic Index coverage score
12.5 / 14 capabilities · 89%
| Integrations & Tool Calling | Full |
|---|---|
|
Tools are implemented and called by agents to invoke external APIs and services, a tool node places a tool inside an agentic workflow, and the ADK documentation carries a tools section covering tool definition, connections to external systems with their own credential and OAuth token handling, an OpenAPI plugin, and MCP-based workflows. Sourceibm.com/products/watsonx-orchestrate/developersread 2026-09-17 |
|
| Workflow Orchestration | Full |
|
Agentic workflows are assembled from documented node types, including an agent node that hands a step to another agent, branch, parallel, foreach and decision nodes, a script node, a timer node that sets a delay between actions, and a user activity node, and IBM describes building with the Agent Development Kit for pro-code control, Langflow for visual design and Flow Builder for rapid orchestration, with branching logic and long-running flows named on the same page. Sourceibm.com/products/watsonx-orchestrate/developersread 2026-09-17 |
|
| Knowledge Grounding & RAG | Full |
|
Agents use knowledge bases to search and return grounded content. Knowledge bases are defined in YAML or JSON and imported through the ADK CLI, in two forms: built-in, backed by a Milvus instance and populated with documents the customer imports, or external, pointing at the customer's own Milvus or Elasticsearch instance. Sourcedeveloper.watson-orchestrate.ibm.com/knowledge_base/overview.mdread 2026-09-17 |
|
| Human Oversight & Guardrails | Full |
|
Agentic workflows carry a user activity node, an interactive step where the workflow waits for a person to provide or receive information, either one field at a time in conversation or through a form that collects several fields at once, with node owners, input and output schemas, and fields that select approvers from a user list. IBM describes the same capability on the product side as workflows that pause for human validation, capture context and resume automatically with audit-ready traceability, so the approval point sits at the node layer, IBM watsonx Orchestrate ADK documentation, User activity node page, and the developers product page, 2026-09-17 Sourcedeveloper.watson-orchestrate.ibm.com/tools/flows/user_activities_node.mdread 2026-09-17 |
|
| Security, Identity & Governance | Partial |
|
Security and access control is named as a feature from the entry plan, and the Premium plan adds data isolation, a dedicated data space, end-to-end protection and a HIPAA-ready option. Against the published questions of which identity providers and provisioning standards are supported, and whether access control reaches users, teams, tools, knowledge sources and model usage, none of that is documented on the pages read, and no certification or audit report scoped to the product was found on IBM's own estate, so the attestation half of the bar is unmet and this sits at Partial on scope rather than on doubt. Sourceibm.com/products/watsonx-orchestrate/pricingread 2026-09-17 |
|
| Observability & Auditability | Full |
|
Traces visualize the complete path a request takes through an agent, with each observation recording a distinct unit of work, and they are reachable four ways: the orchestrate observability traces CLI command, the ADK Python library, the product interface and the API. Sourcedeveloper.watson-orchestrate.ibm.com/traces/overview.mdread 2026-09-17 |
|
| Memory & State Persistence | Full |
|
The Agentic SDK ships memory APIs that persist user information and retrieve it during later agent execution, against documented /memories and /memories/search endpoints: add_messages writes, search, list and retrieve read, and delete and delete_all remove, with memories typed (for example preference), tagged with metadata, scoped by agent and run, and carrying a sensitivity classification. Sourcedeveloper.watson-orchestrate.ibm.com/sdk/memory.mdread 2026-09-17 |
|
| Deployment & Data Residency | Full |
|
IBM states the product runs as a fully managed service on IBM Cloud or on AWS, or in the customer's own on-premises environment, and the ADK documentation carries separate pages for on-premises compatibility and for an air-gapped environment. Sourceibm.com/products/watsonx-orchestrate/pricingread 2026-09-17 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
An agents and tools catalog is a named feature of every plan, with ready-to-use domain agents added at the Standard plan, and the product API exposes catalog management operations for listing the applications and skills available from it. Partner-built agents are sold beside it as separately metered add-ons, each a complete agent a buyer adopts on its own, among them a web research agent, a sales intelligence agent, a messaging agent and a document question-answering agent, IBM watsonx Orchestrate pricing page and ADK API documentation, catalog management operations, 2026-09-17 Sourceibm.com/products/watsonx-orchestrate/pricingread 2026-09-17 |
|
| Triggers & Channel Coverage | Full |
|
Agentic workflows and agents can be scheduled to run automatically at specific times or intervals once schedulable is set on the workflow or agent, with the platform handling cron patterns, timezones and execution limits and the schedule running without manual intervention, and separately each channel a buyer configures (web chat, Twilio WhatsApp, Twilio SMS, Slack, Genesys Bot Connector, Facebook Messenger, Microsoft Teams) provides its own event URL to receive messages, IBM watsonx Orchestrate ADK documentation, Scheduling overview and Channels overview pages, 2026-09-17 Sourcedeveloper.watson-orchestrate.ibm.com/scheduling/overview.mdread 2026-09-17 |
|
| Model Flexibility & Routing | Full |
|
Models from multiple providers are imported as virtual models and assigned to an individual agent, and model policies coordinate several of them with a load-balance, fallback or single strategy, weights per target, and retry on named HTTP status codes; a policy is assigned to an agent the same way a single model is. IBM describes the same surface on the product side as an AI Gateway for integrating multi-provider models with configurable routing policies. Model policies are published as public preview, while per-agent model assignment is not, IBM watsonx Orchestrate ADK documentation, Configuring model policies page, and the developers product page, 2026-09-17 Sourcedeveloper.watson-orchestrate.ibm.com/llm/model_policies.mdread 2026-09-17 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
The platform publishes its own developer documentation site carrying an Agent Development Kit with a Python SDK and an orchestrate command line (agents import, evaluations, observability traces, models policy), a machine-readable documentation index, and a published OpenAPI specification for the server API alongside per-operation reference pages for agents, flows, connections, catalog management, evaluation and monitoring. Sourcedeveloper.watson-orchestrate.ibm.com/llms.txtread 2026-09-17 |
|
| Testing, Debugging & Optimization | Full |
|
The ADK ships an evaluation framework that assesses agent behavior by simulating user interactions and comparing the resulting trajectories step by step against reference data the customer generates, records or writes by hand, logging every deviation from the expected path and returning a metrics summary saved as CSV. The documented pre-deployment cycle is develop, evaluate, analyze, improve, with an analyze command for reading results and a rubric; evaluations run against the local Developer Edition and against draft environments of SaaS and on-premises instances, IBM watsonx Orchestrate ADK documentation, Evaluation overview and Evaluating pages, 2026-09-17 Sourcedeveloper.watson-orchestrate.ibm.com/evaluate/overview.mdread 2026-09-17 |
|
| Browser & Computer Use | Unable to verify |
|
No browser, desktop or remote computer control appears in the platform's own documentation: agents act through tools that call APIs and services, connections to enterprise systems, MCP and the channels a buyer configures, so the question of what happens when an interface changes does not arise for the platform itself. Sourceibm.com/products/watsonx-orchestrate/pricingread 2026-09-17 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Unable to verify against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Recent platform changes
Agents built with watsonx Orchestrate's ADK 2.18.0 can now shortlist tools, retrieving only the ones relevant to the current message before the model plans. IBM says this lowers token use and improves planning for agents with many tools.
Bears on: Agent capability
View sourceIBM's mid month watsonx Orchestrate update adds AI Gateway connections for Gemini and Azure AI Foundry, and agent identity configuration through IBM Verify and Microsoft Entra, along with new content guardrails.
Bears on: Integrations
View sourceIBM announced the general availability of the AgentOps Agent for watsonx Orchestrate, alongside new features for cross-platform agent discovery and custom evaluation. These additions allow organizations to discover agents across different platforms and evaluate their performance using tailored metrics.
Bears on: Observability / auditability
View sourcePricing
Essentials from $530/mo · Standard from $6,360/mo · Premium contact us · 30-day free trial
hybrid
Included quota
Essentials ($530/mo): agent building and orchestration, workflow and document processing, agents and tools catalog, custom tool creation, core AI and LLMs, security and access control; 4,000 monthly active users, up to 200,000 messages a month, up to 10 GB of documents for agent knowledge, channels for web chat, voice, Slack and Teams. Standard ($6,360/mo): the same plus higher usage limits, document volumes, message throughput and concurrency, and ready-to-use domain agents; 40,000 monthly active users, up to 2,000,000 messages a month. Premium (quoted): adds data isolation, a dedicated data space, end-to-end protection, a HIPAA-ready option and enterprise-grade scale.
What is public
IBM watsonx Orchestrate publishes three plans on its own pricing page. Trial: 30-day free access, no permanent free tier. Essentials: starting at $530.00 a month, covering agent building and orchestration, workflow and document processing, the agents and tools catalog, custom tool creation, core AI and LLMs, and security and access control. Standard: starting at $6,360.00 a month, adding higher usage limits, document volumes, message throughput, concurrency and ready-to-use domain agents. Premium: contact us, adding data isolation, a dedicated data space, end-to-end protection, a HIPAA-ready option and enterprise-grade scale. Plan capacity is published as monthly active users (4,000 and 40,000), messages a month (200,000 and 2,000,000) and 10 GB of documents for agent knowledge. Deployment is stated as managed on IBM Cloud or AWS, or in the customer's own on-premises environment.
Billing mechanics
Tiered subscription with two published starting prices (Essentials $530/mo, Standard $6,360/mo) and a quoted Premium plan, bought either self-serve from IBM.com, on AWS, or through sales. Capacity is set per plan by monthly active users, messages per month and document volume for agent knowledge. Optional add-ons, including partner-built agents, are metered separately in API calls, compute credits or messages. A limited eCommerce pilot sells selected partner-built agents bundled with a subscription.
Cost watchouts
Plan caps are the thing to model: monthly active users and messages per month are capped per plan, with documents for agent knowledge capped at 10 GB on both published plans, so growth past a cap means the next plan or an add-on. Partner agents are sold as separately metered add-ons with their own quotas. Premium, which carries data isolation and the HIPAA-ready option, is quoted rather than published.
Variable cost rationale
Cost scales with tier + agent workload volume (orchestration steps, model inference RUs, MAUs, compute); heavier multi-agent automation = higher underlying watsonx consumption.
Additional watchouts
Entry price is material for a small team, capacity is capped per plan rather than pooled, and the plan carrying data isolation and the HIPAA-ready option is quote-only.
Overage / add-ons
Monthly active users and messages per month are capped per plan. Add-ons are sold with their own quotas, including partner agents metered in API calls, compute credits or messages.
Sales call required
No, self serve available
Lowest paid plan
Essentials
Commercial notes
IBM's agentic control plane, sold in three plans with published entry pricing for the two lower plans, self-serve from IBM.com or on AWS, with a 30-day free trial and no permanent free tier.
Key ambiguities
Both published plans carry a starting price, which IBM states is indicative and may vary by country. Premium is quoted. The pricing page was still advertising a 30 percent discount on a first annual Standard subscription with an offer end date of 15 April 2026 when read on 17 September 2026.
Cancellation / refund
30-day free trial (no credit card); Essentials/Standard/Premium via subscription; marketplace self-checkout available for some plans.
Support SLA / resale
Bought self-serve from IBM.com, on AWS with cloud credits, or through IBM sales and partners; agents and tools catalog carries IBM and partner-built agents; managed on IBM Cloud or AWS, or on premises, with an air-gapped environment documented in the ADK.
Missing data
Premium pricing is quoted, not published. Per-usage metering beneath the plan caps is not fully published at the Orchestrate layer.
Related vendors
- 11th Estate — Agentic platform whose AI engine scans markets, matches an…
- Adopt AI — AI CPA firm whose agents run reconciliations, close, AP and AR, and…
- Aera Technology — Decision intelligence platform where an always on agent executes…
- Akro AI — On premise operational intelligence that automates document heavy…
- alfred_ — AI executive assistant that triages the inbox overnight and scores…
- Alloy.ai — Commerce intelligence system for consumer brands that unifies four…
Alternatives to IBM watsonx Orchestrate
The closest documented capability profiles to IBM watsonx Orchestrate among enterprise operations agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Boomi12.5 / 14Fuller documented coverage on Security, Identity & Governance
- Fabrix.ai12.5 / 14Fuller documented coverage on Security, Identity & Governance
- Harvey12.5 / 14Fuller documented coverage on Security, Identity & Governance
- Salesforce12.5 / 14Fuller documented coverage on Security, Identity & Governance
- Appian14.0 / 14Adds documented Browser & Computer Use
- Atlassian12.0 / 14Fuller documented coverage on Security, Identity & Governance
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded