Beam AI

Integrations is one of seven items in the workspace navigation, where a team browses available integrations, adds a connection, reviews existing connections or starts a custom integration. An integration is bound to an individual Flow node from the sidebar's Parameters > Integrations section, so one workflow can act across several systems at different points.

Integration connectors are the tools an agent uses inside workflow steps to take action in external systems, such as sending an email, creating a record, or fetching data from apps like Gmail, Salesforce, Airtable or Slack, and that includes writing records into Salesforce. A browsable catalog is published at beam.ai/integrations, and custom integrations let a customer build an integration for an internal or proprietary platform, so the choice is not limited to what Beam ships. Beam does not publish an integration count.

Sourcedocs.beam.ai core-concepts Tools and integrations, docs.beam.ai integrations and custom-integrations, beam.ai/platform/ai-agents FAQread 2026-09-13

Customers author each workflow as a Flow, a configured graph of nodes and edges that an agent uses to carry out work. The Flow canvas starts with a trigger and can contain tool nodes and LLM based condition nodes, so branches are decided by a model rather than a fixed rule. A loop node iterates over an array or runs a fixed number of passes inside a Flow to process multi-item payloads sequentially, and agents can call other agents for complex workflow orchestration and specialized task delegation. Flows follow a release lifecycle.

The canvas header carries Flow state, Test mode and Publish controls, a builder can run a draft, publish it and review version history, and the API keeps both an active published graph and a draft graph. Builders choose AI, integration, controls or utilities steps and can mix fixed steps with agentic ones in the same workflow, so deterministic steps execute reliably, AI reasoning handles the exceptions, and the builder decides which steps need which approach.

Sourcedocs.beam.ai core-concepts Flows, multi-agent-collaboration, loop-node, publishing-deployment; beam.ai/platform/agentic-automationread 2026-09-13

Each agent has a knowledge store that Beam calls Memory. A builder adds documents from Configuration > Memory > Add data, then uses Memory fill when a node needs knowledge from those files as an input. Memory fill is one of seven variable fill methods (Auto, Prompt, Linked, Static, User, Memory, Attachment), so grounding a step in the documents is a choice the builder makes per input.

The documents persist in a module attached to the agent, and the API manages them with Upload Context File, Download Context File and Delete Context File endpoints, while setup accepts up to ten context files across txt, csv, pdf, xls, xlsx, docx, doc, ppt, pptx and images.

Beam describes its approach as agentic RAG, in which the agent retrieves the most relevant information from databases or knowledge sources, injects that context into the reasoning step, and then uses tools to act on grounded data, connecting agents to existing databases for more accurate, context rich outputs. Beam also markets a Database, Memory & RAG product. Beam does not describe its index, embedding model, chunking strategy, ranking method or any citation back to the source file.

Sourcedocs.beam.ai core-concepts Memory and Variables, docs.beam.ai API agent-context-files and agent-setup, beam.ai/platform/ai-agents FAQread 2026-09-13

Beam ships a review and approve gate. Automation modes configure human-in-the-loop workflows with three interruption types, consent approvals, input requests and failure handling, set per workflow. Autonomy is a setting with three values. Fully autonomous runs end to end without human intervention, human in the loop pauses at designated checkpoints for review, and hybrid runs autonomously with selective oversight, so the customer places the checkpoints and can tier risk across the steps of one workflow.

All pending approvals route to a centralized Inbox in the workspace navigation, beside Skills, Projects, Tasks, Agent templates, Integrations and Views. The gate is enforced in the execution engine as well as the interface. Approve Task Execution gives consent for a task to continue when approval is required, Reject Task Execution marks it rejected and stops execution, and Submit User Input provides input for a task that needs a person to continue.

Beam suggests the gate for sensitive decisions like high-value invoices, customer escalations and contract modifications, where agents prepare everything and humans make the final call, and for risky steps like sending external emails, updating customer records or approving financial actions, while low risk steps run autonomously. Role based access also governs who may run an agent, and an audit trail records what happened.

Sourcebeam.ai/platform/agentic-automation, docs.beam.ai automation-modes, docs.beam.ai API agent-task approve/reject/submit-user-inputread 2026-09-13

Beam is SOC 2 Type II compliant, with the report available on request by email, and has GDPR and HIPAA sections and a trust center at security.beam.ai. Workspace roles decide which agents each teammate can use and what they can do with them, and Azure SSO handles enterprise authentication.

An RBAC release lets administrators assign built in or custom roles governing who can view, create, run or delete specific agents, so running an agent, which spends money and touches production, can be limited per agent. Supporting controls include AES-256 and TLS encryption, annual employee security training, 30 day backup retention, breach protocols, and third party penetration testing and domain vulnerability scanning. Beam does not publish an ISO 27001 certification.

Sourcebeam.ai/legal/security, docs.beam.ai roles-and-permissions and team-management-ssoread 2026-09-13

A centralized analytics dashboard reports agent performance, task completion rates, evaluation scores and user feedback, and Analytics is also one of the six areas of each agent, beside Chat, Tasks, Flow, Learning and Configuration, so the view exists per agent and per workspace.

Each run is a task, one execution of an agent, and the Tasks page lets a user review task records, switch between tasks and tests, filter and search by task name, and open individual tasks for execution details. Debug tools go down to the step, inspecting a task execution and letting a builder configure a Flow node with sample input before testing its output.

Monitoring sends task status notifications for an agent on a chosen schedule and channel, for chosen task statuses, with or without test tasks. All of it is available through the API, where Get Agent Analytics returns success rates, performance metrics and usage statistics, Get Latest Task Executions serves monitoring and debugging, Get Task Details returns execution status and results, and Get Task Updates (SSE) streams status changes in real time. Governance includes RBAC, approvals and audit logs, with a full audit trail for compliance.

Sourcedocs.beam.ai overview-analytics, monitoring, task-executions, debug-tools; API agent-task endpoints; beam.ai/platform/agentic-automationread 2026-09-13

Beam's feature called Memory is a document store. A user adds documents from Configuration > Memory > Add data, and Memory fill pulls knowledge from those files into a node as an input, beside Auto, Prompt, Linked, Static, User and Attachment fill. Beam also says agents become better at understanding and executing a task the more people interact with them, with each executed task a step toward augmenting intelligence further.

A marketing FAQ on agent memory gives a hedged, general definition of the concept, a context layer that lets agents carry state across steps and, when needed, across sessions, and no product feature matches it. Beam does not describe a memory scope, lifetime, expiry or purge path, or any way to read and write accumulated state, and nothing shows an agent recalling a prior run's outcome. Google Drive and SharePoint connections that would keep an agent's Memory in sync with repository folders are planned and not yet available.

Sourcedocs.beam.ai core-concepts Memory and Variables, beam.ai/platform/ai-agents Learning agentread 2026-09-13

Customers can deploy Beam on premises or in the cloud, with custom solutions as a third option, and Beam says the platform fits into a customer's ecosystem regardless of preferred infrastructure. On premises means the platform itself runs in the customer's own environment, not a connector reaching into it. Beam describes its hosting as enterprise grade, EU-hosted and SOC 2, with a commitment to keep data localized. Beam sets out the trade off between the two.

Cloud deployment optimizes for fast rollout, elasticity and simpler operations, while on premises deployment optimizes for tighter data residency requirements, internal network access and stricter governance for regulated environments. Beam publishes no region list, self service region selector or residency configuration, and no installation or architecture guide for the on premises path.

Sourcebeam.ai/platform/ai-agents Customizable deployment and Security & privacy, FAQ cloud versus on premisesread 2026-09-13

Agent templates is a browsable catalog in the workspace navigation, where a team browses, filters and creates agents from the templates available in its workspace, and a public gallery sits at beam.ai/agents. A template arrives as a working agent with a configured Flow and wired integrations, which the team reviews, tests and publishes when it is ready, and Beam describes them as pre-trained AI agents to start in seconds from a pre-made selection.

Templates such as a Patient Intake Scheduler, a Social Media Content agent and an Order Tracking agent are each recommended for a different industry, and each works on its own. Workspace skills add a second layer of reusable capabilities that a team can browse, search and create, with a public page at beam.ai/skills and an API endpoint that imports a skill into an agent from a zip file. Skills are components that sit inside agents. Beam's industry landing pages for Banking, Healthcare, Insurance, BPO, Debt Collection and Property Management are marketing segments, not packaged agents.

Sourcedocs.beam.ai template-gallery and quickstart, docs.beam.ai workspace-skills, beam.ai/platform/ai-agents Templatesread 2026-09-13

Every Flow begins with a trigger node, where a builder reviews existing trigger cards or adds a trigger, and the configuration depends on the trigger type. Several trigger cards can attach to one Flow, and the customer configures and tests the trigger node before publishing, with a Gmail trigger editor as the worked example. Webhooks give any external system an inbound event to call.

A trigger defines when an agent starts and what input data passes into the workflow, and that input becomes the task query every node can read. Beam lists intelligent triggers as a capability alongside tools and integrations, and the public API's Create Agent Task queues an execution as a separate programmatic path.

Sourcedocs.beam.ai core-concepts Triggers and triggers-webhooks reference, beam.ai/platform/ai-agents Integrations tools and triggersread 2026-09-13

Model is one of the parameters on every tool node, so the builder chooses it per step and it can differ between steps of the same workflow. The node sidebar shows Parameters, Settings and Evaluation, and Parameters holds the input variables, model, output variables and integrations. The API publishes a Get Preferred Models endpoint in its custom tool section.

Beam calls the platform model-agnostic and API-first, so workflows can evolve without lock-in, and says teams commonly pair it with OpenAI, DeepSeek or other LLMs. An LLM directory at beam.ai/llm has pages per provider. Beam does not list the models a builder can select.

Sourcedocs.beam.ai core-concepts Flows, docs.beam.ai API get-preferred-models, beam.ai/platform/ai-agents FAQread 2026-09-13

Beam publishes a public REST API that builds agents as well as reading them, at docs.beam.ai/08-reference/api, with a base URL of api.beamstudio.ai, API key authentication through an x-api-key header plus a current-workspace-id header, an Authentication guide and worked code examples. An OpenAPI specification is published at /openapi-final.json. The endpoints author agents programmatically.

Create a complete agent graph builds a new agent with its full graph of nodes, edges and tool configurations as both published and draft graphs, Update agent and its draft graph replaces all nodes and edges, Add node to graph and Add edge to graph compose the flow, Update prompt of the tool and Update input/output parameters reach inside a node, and Publish agent graph ships it. A caller outside the product can author, wire, configure, test and publish an agent without opening the interface.

Runtime is exposed too, through Create Agent Task, List Agent Tasks, Get Task Details, Retry Task Execution, Approve and Reject Task Execution, Submit User Input, and Get Task Updates (SSE) for a real time stream of execution status. Beam also offers an MCP connection, imports skills into an agent from a zip file, and lets customers build custom integrations for internal or proprietary platforms.

It calls the platform model-agnostic and API-first, so workflows can evolve without lock-in, and lets teams embed agents through components and APIs, manage tenants and usage, and enforce policies centrally.

Sourcedocs.beam.ai API reference, authentication, MCP connection and openapi-final.json; beam.ai/platform/ai-agentsread 2026-09-13

Test mode lets a builder run an agent flow with data from past tasks in a safe environment before publishing, so a change is tried on real production tasks against an unpublished draft, and the draft and published graph model builds that step into the release. Rerunning tasks supports debugging, testing workflow changes, backtesting prompt improvements and validating agent performance.

The evaluation framework validates agent accuracy with evaluation criteria, auto-run retries and analytics monitoring, and the analytics dashboard shows evaluation scores beside completion rates. Each Flow node has Parameters, Settings and Evaluation tabs, so criteria attach to individual tool nodes, and auto run retries mean a failed criterion changes runtime behavior rather than only reporting. The improvement loop is programmatic as well.

Rate Task Output submits a rating for a task's output to improve agent performance, Optimize Tool tunes a tool's performance and configuration based on usage patterns and feedback, with Get Tool Optimization Status to track it, and Pro learning tools test, tune and optimize an agent's tool outputs. Beam's Learning Hub aggregates agent feedback and self checks, spots recurring failures and drafts the instruction fix for a person to approve.

Sourcedocs.beam.ai evaluation-framework, test-datasets, rerunning-tasks, optimize-outputs; API rate-task-output and optimize-toolread 2026-09-13

A Flow can contain AI, integration, controls or utilities steps, and none of them drives a browser session, a screen, or a remote or local computer. Agents reach other systems through integration connectors and APIs on an API first platform, taking actions such as sending an email, creating a record, or fetching data from apps like Gmail, Salesforce, Airtable or Slack, so if Salesforce redesigned its interface a Beam Flow would be unaffected.

Beam positions itself as the successor to robotic process automation, and an FAQ lists RPA agents that capture invoices, post orders, reconcile payments and validate forms among examples of AI agents, but that passage describes the general category, and no product page says a Beam agent drives a screen. Beam's own blog singles out a rival platform's ability to operate GUIs for apps without APIs as what stands out about that rival.

Sourcedocs.beam.ai core-concepts and tools-integrations, beam.ai/platform/ai-agentsread 2026-09-13