Beam AI

Full CONFIRMED and re-based off the rubric. THE INTEGRATION LAYER IS A FIRST-CLASS WORKSPACE MODULE, NOT A LOGO WALL. Integrations is one of seven items in the workspace navigation beside Skills, Projects, Inbox, Tasks, Agent templates and Views, and the docs describe operating it: USE THE WORKSPACE INTEGRATIONS PAGE TO BROWSE AVAILABLE INTEGRATIONS, ADD A CONNECTION, REVIEW EXISTING CONNECTIONS, OR START A CUSTOM INTEGRATION. Browse, connect, review and extend is a managed connector estate with a lifecycle.

BINDING IS PER STEP: an integration is attached to an individual Flow node from the sidebar's Parameters > Integrations section, so a single workflow can act across several systems at different points.

NAMED COUNTERPARTIES ARE PUBLISHED: INTEGRATIONS, VIA INTEGRATION CONNECTORS, ARE THE TOOLS THE AGENT USES INSIDE WORKFLOW STEPS TO TAKE ACTION IN EXTERNAL SYSTEMS, FOR EXAMPLE SENDING AN EMAIL, CREATING A RECORD, OR FETCHING DATA FROM APPS LIKE GMAIL, SALESFORCE, AIRTABLE, OR SLACK, and a browsable catalogue is published at beam.ai/integrations, linked from every footer and from a SEE ALL INTEGRATIONS control. THE THRESHOLD CLAUSE IS ACTION, NOT READ, and the verbs are the vendor's own — sending, creating, fetching.

Writing a record into Salesforce is an authenticated write into a system of record, which is what separates this from archimetis's read-only Partial one record earlier. CUSTOM INTEGRATIONS CLOSE THE GAP THE CATALOGUE LEAVES: a documented path to BUILD AN INTEGRATION FOR AN INTERNAL OR PROPRIETARY PLATFORM FROM THE WORKSPACE INTEGRATIONS PAGE means the estate is not bounded by what Beam shipped. A CLAIM CARRIED IN THE INDEX'S OWN COMPARISON PROSE IS NOT RELIED ON: two published pairs assert 1000+ INTEGRATIONS, and no first-party page read this pass states a count.

The grade rests on the documented module, the named counterparties and the custom path, not on the number, and the pair prose should be checked against a first-party source at the sublane close. CONFIDENCE 0.8.

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

Full CONFIRMED and re-based off the rubric, AND THIS IS A CUSTOMER-COMPOSABLE WORKFLOW LAYER OF THE KIND archimetis LACKED ONE RECORD AGO — the contrast is the clearest in the sublane so far. THE ARTIFACT IS AN EXPLICIT GRAPH THE BUYER AUTHORS: A FLOW IS THE CONFIGURED GRAPH THAT AN AGENT USES TO CARRY OUT WORK. THE CURRENT FLOW CANVAS STARTS WITH A TRIGGER AND CAN CONTAIN TOOL AND LLM-BASED CONDITION NODES.

Nodes, edges, and conditions evaluated by a model rather than a static rule. CONTROL FLOW IS REAL, NOT LINEAR: LLM-based Condition nodes branch, and a documented LOOP NODE will ITERATE OVER AN ARRAY OR RUN A FIXED NUMBER OF PASSES INSIDE A FLOW TO PROCESS MULTI-ITEM PAYLOADS SEQUENTIALLY. Branching plus iteration is orchestration in the strict sense.

MULTI-AGENT DELEGATION IS DOCUMENTED SEPARATELY: ENABLE AGENTS TO CALL OTHER AGENTS FOR COMPLEX WORKFLOW ORCHESTRATION AND SPECIALIZED TASK DELEGATION. An agent invoking another agent is the multi-actor case, evidenced rather than asserted.

FLOWS ARE VERSIONED AND HAVE A RELEASE LIFECYCLE, which is what makes them durable rather than per-run assembly: the canvas header carries Flow state, Test mode and Publish controls, and the PUBLISHING & DEPLOYMENT page covers running a draft, publishing it, and reviewing VERSION HISTORY. The API mirrors it exactly, maintaining BOTH AN ACTIVE/PUBLISHED GRAPH AND A DRAFT GRAPH. Draft, test, publish, version is a software release model applied to a workflow.

STEP TYPES ARE TYPED AND MIXED: builders choose AI, INTEGRATION, CONTROLS, OR UTILITIES steps, and the vendor's positioning is that both modes coexist — BEAM COMBINES FIXED FLOWS WITH AGENTIC APPROACHES IN THE SAME WORKFLOW. DETERMINISTIC STEPS EXECUTE RELIABLY... AI REASONING HANDLES THE EXCEPTIONS. YOU DESIGN WHICH STEPS NEED WHICH APPROACH. Letting the builder decide per step which parts are deterministic is a more governed orchestration model than either extreme. NOT COUNTED HERE: the Trigger node that starts a Flow is Trig; the integrations bound to nodes are Int; approval checkpoints are HITL. CONFIDENCE 0.85.

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

Full CONFIRMED and re-based off the rubric, AND IT TAKES THE FACT THE APRIL GRID SPENT ON Mem. THE MAINTAINED STRUCTURE IS PER-AGENT AND HAS A DOCUMENTED CONFIGURATION PATH: EACH AGENT HAS A MEMORY MODULE. ADD DOCUMENTS FROM CONFIGURATION > MEMORY > ADD DATA. Beam calls it Memory and it is a knowledge corpus — the vendor's own sentence continues THEN USE MEMORY FILL WHEN A NODE NEEDS KNOWLEDGE FROM THOSE FILES AS AN INPUT. Knowledge from files, injected into a step. That is this axis, and it is refused on Mem accordingly.

THE RETRIEVAL METHOD IS A FIRST-CLASS, NAMED MECHANISM. Memory fill is one of seven documented variable fill methods (Auto, Prompt, Linked, Static, User, Memory, Attachment), so grounding a step in the corpus is a configuration choice a builder makes explicitly per input rather than an implicit behaviour.

THE CORPUS HAS A FULL LIFECYCLE API, which is the strongest evidence that it is a managed structure rather than a prompt-stuffing convenience: Upload Context File, Download Context File and Delete Context File are published endpoints, and setup accepts up to ten context files across txt, csv, pdf, xls, xlsx, docx, doc, ppt, pptx and images. Upload, list, download and delete is a document store the vendor operates on the customer's behalf.

THE VENDOR DESCRIBES THE RETRIEVAL PATTERN EXPLICITLY: agentic RAG means the agent RETRIEVES THE MOST RELEVANT INFORMATION FROM YOUR DATABASES OR KNOWLEDGE SOURCES, INJECTS THAT CONTEXT INTO THE REASONING STEP, AND THEN USES TOOLS TO EXECUTE ACTIONS BASED ON GROUNDED DATA, positioned as CONNECTING AI AGENTS TO EXISTING DATABASES TO PRODUCE MORE ACCURATE, CONTEXT RICH OUTPUTS. A dedicated DATABASE, MEMORY & RAG product page is carried in every site footer.

THIS SITS ON THE MAINTAINED-STRUCTURE SIDE OF THE LINE THIS LANE WATCHES, not the per-run assembly side: the documents persist in a module attached to the agent and are selected per input, rather than being gathered fresh each execution. WHY CONFIDENCE IS 0.75: no index, embedding model, chunking strategy or ranking method is disclosed, and no citation or source-attribution behaviour is documented — nothing says an output points back to the file it came from, which shortwave, writer and alfred_ each showed. The structure clears the bar; the provenance layer is undocumented. The DATABASE, MEMORY & RAG page was not opened within budget and is the door that would settle it. CONFIDENCE 0.75.

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

Full CONFIRMED and re-based off the rubric, and this is a shipped review-and-approve gate rather than a reserved decision — the distinction the standing rule turns on. THE GATE HAS ITS OWN DOCUMENTATION PAGE: AUTOMATION MODES covers CONFIGURING HUMAN-IN-THE-LOOP WORKFLOWS WITH CONSENT APPROVALS, INPUT REQUESTS, AND FAILURE HANDLING FOR CONTROLLED AGENT AUTONOMY. Three distinct interruption types — consent, input, failure — configured per workflow.

AUTONOMY IS A SETTING WITH THREE NAMED VALUES: FULLY AUTONOMOUS (END-TO-END WITHOUT HUMAN INTERVENTION), HUMAN-IN-THE-LOOP (PAUSE AT DESIGNATED CHECKPOINTS FOR REVIEW), AND HYBRID (AUTONOMOUS WITH SELECTIVE OVERSIGHT). Designated checkpoints mean the buyer places the gate, and the hybrid mode means risk can be tiered across steps of one workflow rather than all-or-nothing.

THE APPROVAL SURFACE IS A REAL PLACE IN THE PRODUCT, WHICH IS WHAT MOST RECORDS ON THIS AXIS CANNOT SHOW: ALL PENDING APPROVALS ROUTE TO A CENTRALIZED INBOX, and Inbox appears in the documented workspace navigation beside Skills, Projects, Tasks, Agent templates, Integrations and Views. A queue with a named home is a shipped mechanism, not a posture.

THE API CONFIRMS IT IS ENFORCED IN THE EXECUTION ENGINE RATHER THAN THE UI: APPROVE TASK EXECUTION provides USER CONSENT FOR A TASK TO CONTINUE EXECUTION WHEN APPROVAL IS REQUIRED; REJECT TASK EXECUTION marks it rejected and STOPS EXECUTION; SUBMIT USER INPUT provides input FOR A TASK THAT REQUIRES HUMAN INTERACTION TO CONTINUE. A task that halts pending consent and resumes or dies on the answer is a genuine gate in the runtime.

THE VENDOR SCOPES IT TO THE RIGHT ACTIONS: FOR SENSITIVE DECISIONS LIKE HIGH-VALUE INVOICES, CUSTOMER ESCALATIONS, CONTRACT MODIFICATIONS, AGENTS PREPARE EVERYTHING, HUMANS MAKE THE FINAL CALL, with the design guidance to GATE RISKY STEPS LIKE SENDING EXTERNAL EMAILS, UPDATING CUSTOMER RECORDS, OR APPROVING FINANCIAL ACTIONS, WHILE LETTING LOW RISK STEPS RUN AUTONOMOUSLY. PERMISSIONING IS A SECOND, INDEPENDENT LAYER and is spent on Sec, not here: RBAC governs who may run an agent at all. NOT COUNTED: FULL AUDIT TRAIL FOR COMPLIANCE, which is Obs. CONFIDENCE 0.85.

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

Full CONFIRMED and re-based off a rubric onto first-party pages. The April cell read only AGENT FEATURES RESEARCH REPORT + JSON FEATURE RUBRIC at 0.85 confidence with no URL, so the grade was right and rested on nothing. BOTH HALVES OF THE CONJUNCTION ARE NOW EVIDENCED SEPARATELY.

ATTESTATION: the security page states BEAM IS AN SOC 2 TYPE II COMPLIANT COMPANY, with the report available on request — ENQUIRE ABOUT OUR CERTIFICATION BY EMAILING US — alongside GDPR and HIPAA sections, and a TRUST CENTER at security.beam.ai linked from the page body and from every site footer. Naming the standard, the Type, and an access path to the report is the asserted form, not the ALIGNED WITH or READY hedge that sits at Partial under Q50.

ACCESS: the docs carry a ROLES AND PERMISSIONS page — CREATE WORKSPACE ROLES THAT DECIDE WHICH AGENTS EACH TEAMMATE CAN USE AND WHAT THEY CAN DO WITH THEM — and a dedicated AZURE SSO SETUP page for enterprise authentication. The index's own change log records the RBAC release on 2026-09-01: administrators assign built-in or custom roles governing who can VIEW, CREATE, RUN, OR DELETE specific agents.

Per-agent authorisation over the RUN verb is a customer-facing control surface of unusual precision, since run is the verb that spends money and touches production. SUPPORTING CONTROLS, NOT COUNTED AS EITHER HALF: AES-256 and TLS encryption, annual employee security training, 30-day backup retention, documented breach protocols, and third-party penetration testing and domain vulnerability scanning.

ONE CLAIM IN THE RECORD'S OWN PROSE IS NOT SUPPORTED AND IS STRUCK: the longDescription asserted CERTIFICATIONS INCLUDING SOC 2 TYPE II, ISO 27001, AND GDPR. ISO 27001 APPEARS NOWHERE on the security page, the platform pages or the docs. SOC 2 and GDPR are real; ISO 27001 was a rubric invention and publishing it would have been a false fact about a named company. NOT COUNTED HERE BY THE ONE-FACT RULE: audit trails are Obs; EU hosting and on-premise deployment are Dep. CONFIDENCE 0.85.

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

Full CONFIRMED and re-based off the rubric, AND THE OBJECT IS CORRECT — every surface below reads on Beam's own agents, not on a customer's operational estate, which is the defect this axis collects. OBSERVABILITY IS A NAMED SECTION OF THE DOCUMENTATION, not a scattered claim: docs.beam.ai carries an 04-OBSERVABILITY-ANALYTICS tree with four pages.

THE DASHBOARD IS CENTRAL AND ITS METRICS ARE AGENT METRICS: MONITOR AGENT PERFORMANCE, TASK COMPLETION RATES, EVALUATION SCORES, AND USER FEEDBACK THROUGH THE CENTRALIZED ANALYTICS DASHBOARD. Completion rate, evaluation score and feedback are three different readings of how the agent is doing. Analytics is also one of the six areas of an individual agent alongside Chat, Tasks, Flow, Learning and Configuration, so the view exists per agent as well as per workspace.

THE EXECUTION RECORD IS INSPECTABLE DOWN TO THE RUN: a task is ONE EXECUTION OF AN AGENT, and the Tasks page lets a user REVIEW TASK RECORDS, SWITCH BETWEEN TASKS AND TESTS, FILTER THE LIST, SEARCH BY TASK NAME, and open INDIVIDUAL TASKS FOR EXECUTION DETAILS. Filtering a task history and opening one execution is exactly the run-history view archimetis and apprentice-io could not produce.

DEBUGGING GOES BELOW THE RUN TO THE STEP: DEBUG TOOLS INSPECT AGENT TASK EXECUTION AND CONFIGURE A FLOW NODE WITH SAMPLE INPUT BEFORE TESTING ITS OUTPUT. ALERTING IS CONFIGURABLE: the Monitoring page covers TASK-STATUS NOTIFICATIONS FOR AN AGENT BY CHOOSING A SCHEDULE, CHANNEL, TASK STATUSES, AND WHETHER TO INCLUDE TEST TASKS — a buyer decides which outcomes page them and where.

IT IS ALL AVAILABLE PROGRAMMATICALLY, which rules out a screenshot-only surface: 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.

AUDIT TRAIL IS ASSERTED ON THE PRODUCT PAGE and is spent here rather than on Sec, per the lane brief: FULL AUDIT TRAIL FOR COMPLIANCE, with governance summarised elsewhere as RBAC, APPROVALS, AUDIT LOGS. NOT COUNTED: evaluation scores are read here as an observability metric and their production is graded on Eval. CONFIDENCE 0.85.

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

CORRECTED Full to None, AND THIS IS THE DATABASE-AS-MEMORY SHAPE WITH THE VENDOR'S OWN NAMING BUILT INTO THE TRAP. Beam ships a feature called Memory. It is a document store. The docs are unambiguous: EACH AGENT HAS A MEMORY MODULE. ADD DOCUMENTS FROM CONFIGURATION > MEMORY > ADD DATA, THEN USE MEMORY FILL WHEN A NODE NEEDS KNOWLEDGE FROM THOSE FILES AS AN INPUT. Files a user uploads, retrieved as input to a step, is a corpus and a retrieval method.

The vendor's own wording is WHEN A NODE NEEDS KNOWLEDGE FROM THOSE FILES — knowledge, not state. IT IS SPENT ON Know AND CANNOT BE SPENT AGAIN HERE. Memory fill sits in a list of seven variable fill methods beside Auto, Prompt, Linked, Static, User and Attachment fill, which places it as a way of getting data INTO a node, not a store of what the agent did.

WHAT REMAINS AFTER THE CORPUS IS SUBTRACTED IS ABSORBED LEARNING, None BY STANDING CONVENTION: AGENTS BECOME BETTER AND BETTER IN UNDERSTANDING AND EXECUTING A TASK, AS YOU INTERACT MORE WITH THEM. EACH TASK EXECUTED IS A STEP TOWARDS AUGMENTING INTELLIGENCE EVEN FURTHER. A model improving is not a memory layer. THE ONE PASSAGE THAT READS AS RUN STATE IS A GLOSSARY ENTRY AND IS REFUSED.

An FAQ headed WHAT IS AGENT MEMORY answers with the general definition of the concept — a context layer letting agents CARRY STATE ACROSS STEPS AND, WHEN NEEDED, ACROSS SESSIONS — hedged with IN BEAM'S FRAMING and IT IS ALSO DESCRIBED AS. That is the vendor defining an industry term on a marketing page, not documenting a shipped store, and no product surface anywhere corresponds to it. Where the docs and a glossary disagree about what ships, the docs govern.

TESTED AGAINST THE Q78 GOVERNED RUN STATE RUNG, ZERO OF FOUR: no memory scope, no published lifetime, no expiry or purge path, no read or write surface for accumulated state. THE PARTIAL RUNG IS NOT REACHED either — nothing documents the agent recalling a prior run's outcome. THE ROADMAP LINE CONFIRMS THE STORE IS A DOCUMENT REPOSITORY: GOOGLE DRIVE AND SHAREPOINT CONNECTIONS FOR KEEPING AN AGENT'S MEMORY IN SYNC WITH REPOSITORY FOLDERS ARE PLANNED BUT ARE NOT YET AVAILABLE.

Syncing memory with folders is what you say about a corpus, and future tense is not evidence in any case. WHAT WOULD MOVE THE CELL: a documented store the agent writes execution state to and reads back, with a scope. CONFIDENCE 0.8.

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

Full CONFIRMED and re-based off the rubric, and it clears the bar TWICE OVER — by a named customer environment and by a named region, either of which alone would do. THE CUSTOMER ENVIRONMENT OPTION HAS ITS OWN HEADED SECTION, CUSTOMIZABLE DEPLOYMENT: DEPLOY YOUR SOLUTIONS EITHER ON-PREMISES OR IN THE CLOUD; OUR PLATFORM IS DESIGNED TO SUPPORT YOUR ORGANIZATION IN SMOOTHLY INTEGRATING AGENTS INTO YOUR ECOSYSTEM, REGARDLESS OF YOUR PREFERRED INFRASTRUCTURE, itemised as CUSTOM SOLUTIONS, CLOUD, ON PREMISE.

On-premise means the software runs in the buyer's own environment, which is the strongest form this axis recognises. THE REGION IS NAMED SEPARATELY: the SECURITY & PRIVACY section on the same page itemises ENTERPRISE GRADE, EU-HOSTED, SOC 2, with the surrounding text committing to ENSURING YOUR DATA REMAINS LOCALIZED. A named jurisdiction plus a commitment is a residency claim, not a reliability property.

THE CHOICE IS PRESENTED TO THE BUYER AS A DECISION WITH STATED TRADE-OFFS, which is what a selection surface looks like in prose: CLOUD DEPLOYMENT OPTIMIZES FOR FAST ROLLOUT, ELASTICITY, AND SIMPLER OPERATIONS... ON PREMISES DEPLOYMENT OPTIMIZES FOR TIGHTER DATA RESIDENCY REQUIREMENTS, INTERNAL NETWORK ACCESS, AND STRICTER GOVERNANCE FOR REGULATED ENVIRONMENTS.

BEAM EXPLICITLY SUPPORTS BOTH ON PREMISES AND CLOUD DEPLOYMENT, AND IT HIGHLIGHTS PRIVACY AND SECURITY SIGNALS LIKE DATA LOCALIZATION, EU HOSTED OPTIONS, AND SOC 2 ALONGSIDE THAT FLEXIBILITY. A vendor naming the reason a buyer would pick each option is describing a real choice. THE CONTRAST WITH THE TWO PRECEDING RECORDS IS THE POINT OF RECORDING THIS AT LENGTH. apprentice-io and archimetis both moved to None this session because network topology — PrivateLink, VPC peering, an on-prem CONNECTOR — was mistaken for deployment topology.

Beam is the genuine article: not a connector reaching into the plant, but the platform itself installed in the customer's infrastructure, plus a named hosting jurisdiction. The distinction is now evidenced from both sides within one sublane. WHY CONFIDENCE IS 0.8 RATHER THAN HIGHER: no region LIST, self-service region selector, or documented residency configuration appears, and the on-premise path is described in marketing prose rather than in an installation or architecture doc. The claim is clear and the mechanics are not published. CONFIDENCE 0.8.

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

Full CONFIRMED and re-based off the rubric. THE BROWSABLE CATALOGUE IS THE BAR AND IT IS A NAMED PRODUCT SURFACE: AGENT TEMPLATES sits in the documented workspace navigation, and its reference page covers BROWSE, FILTER, AND CREATE AGENTS FROM THE TEMPLATES AVAILABLE IN YOUR BEAM WORKSPACE. Browse and filter is a catalogue; create-from is selection. A public gallery is published outside the product as well, at beam.ai/agents behind a SEE ALL AGENTS control.

THE SELECTED UNIT DOES WORK WHEN SELECTED, which is the other half: the quickstart is CREATE AN AGENT FROM A TEMPLATE, REVIEW ITS FLOW AND INTEGRATIONS, TEST IT, AND PUBLISH WHEN IT IS READY. A template arrives carrying a configured Flow and wired integrations, so what a buyer adopts is a working agent rather than a blank shape — the vendor's phrasing is START IN SECONDS WITH PRE-TRAINED AI AGENTS and PRE-MADE SELECTION.

THE REMOVAL TEST PASSES CLEANLY, WHICH IS WHERE apprentice-io PASSED AND archimetis FAILED ONE AND TWO RECORDS AGO. The published units are individually named and unrelated — a Patient Intake Scheduler, a Social Media Content agent, an Order Tracking agent are each cited by the vendor as the recommendation for a different industry. Remove the Order Tracking agent and the Patient Intake Scheduler is still a whole product doing unrelated work.

Contrast archimetis, where removing Ammonia subtracted a paragraph about process coverage and nothing a buyer adopts. A SECOND REUSABLE-UNIT LAYER EXISTS AND IS RECORDED RATHER THAN DOUBLE-COUNTED: WORKSPACE SKILLS lets a team BROWSE, SEARCH, AND START CREATING REUSABLE CAPABILITIES, with a public page at beam.ai/skills and an API endpoint importing a skill into an agent from a zip file.

Skills are components inside agents rather than agents themselves, so they corroborate the catalogue pattern without being counted as a second pack. THE LANE BRIEF'S WARNING WAS TESTED AND DOES NOT BITE: a product line is not a pack, and the industry landing pages (Banking, Healthcare, Insurance, BPO, Debt Collection, Property Management) ARE marketing segmentation and are not counted. What is counted is the template gallery, where the units are individually selectable and individually complete. CONFIDENCE 0.8.

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

Full CONFIRMED and re-based off the rubric onto the product documentation. THE TRIGGER IS A STRUCTURAL ELEMENT OF EVERY AGENT, NOT AN OPTIONAL FEATURE: THE FLOW CANVAS BEGINS WITH A TRIGGER NODE. SELECT IT TO REVIEW EXISTING TRIGGER CARDS OR ADD A TRIGGER. THE DETAILED TRIGGER CONFIGURATION DEPENDS ON THE TRIGGER TYPE. An agent cannot exist without one, multiple trigger cards can attach to a single Flow, and configuration varies by type — which means a typed, extensible trigger system rather than a single hook.

THE CUSTOMER CONFIGURES THEM, WHICH IS THE HALF archimetis COULD NOT SHOW ONE RECORD AGO. A dedicated TRIGGERS AND WEBHOOKS reference page documents configuring the trigger node and testing it before publishing, with a worked Gmail trigger editor as the verified example. Webhooks give an inbound event surface any external system can call.

THE VENDOR STATES THE SEMANTICS PRECISELY: TRIGGERS DEFINE WHEN AN AGENT SHOULD START AND WHAT INPUT DATA IS PASSED INTO THE WORKFLOW, IN BEAM'S DOCS THIS BECOMES THE TASK QUERY THAT ALL NODES CAN ACCESS. A trigger that both starts the run and binds its input payload into a variable every node can read is a real invocation mechanism, not a scheduler bolted on the side.

TRIGGERS ARE POSITIONED AS A CORE BUILDING BLOCK ALONGSIDE TOOLS AND INTEGRATIONS: AGENTS LEVERAGE A DYNAMIC COMBINATION OF TOOLS, INTEGRATIONS, AND EXTERNAL TRIGGERS... TO FLAWLESSLY EXECUTE WORKFLOWS, with INTELLIGENT TRIGGERS itemised as a capability. PROGRAMMATIC INVOCATION EXISTS AS WELL, which is a distinct path from event triggers and is recorded rather than double-counted: the public API publishes CREATE AGENT TASK to queue an execution. TESTED AND NOT NEEDED FOR THE GRADE: the /platform/agentic-automation page's autonomy language, which is spent on HITL. CONFIDENCE 0.85.

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

CORRECTED Unresolved to Full, RESOLVING ONE OF THE INDEX'S 62 U CELLS. The April cell carried U at 0.35 confidence on a rubric with no URL — ignorance recorded honestly as ignorance, which was the right code at the time and is now answerable from the vendor's own documentation. THE AXIS TEST IS WHO CHOOSES, AND THE CUSTOMER CHOOSES, PER STEP. The docs describe the Flow canvas node sidebar: WHEN YOU SELECT A TOOL NODE, THE SIDEBAR INCLUDES PARAMETERS, SETTINGS, AND EVALUATION.

PARAMETERS INCLUDES INPUT VARIABLES, MODEL, OUTPUT VARIABLES, AND INTEGRATIONS. Model is a configurable parameter of every tool node, sitting beside inputs, outputs and integrations, so the choice is made by whoever builds the agent and can differ between steps of the same workflow. Per-node selection is a stronger form of customer control than a single account-level setting.

THE API CONFIRMS IT PROGRAMMATICALLY: a GET PREFERRED MODELS endpoint is published under the custom-tool section of the reference, which only exists because a selectable set exists. THE VENDOR STATES THE POSTURE AND NAMES PROVIDERS: THE PLATFORM IS MODEL-AGNOSTIC AND API-FIRST, SO YOU CAN EVOLVE YOUR AGENTIC WORKFLOWS WITHOUT LOCK-IN, and BRING THE STACK YOU KNOW. TEAMS COMMONLY PAIR BEAM WITH OPENAI, DEEPSEEK, OR OTHER LLMS.

A published LLM directory exists on the marketing estate at beam.ai/llm with per-provider pages. Two named providers plus a directory is disclosure, not a bare POWERED BY AI. THIS IS THE OPPOSITE OF THE SUBLANE PATTERN LOGGED ON THE FOUR RECORDS BEFORE IT — algebra-ai, apaleo, apprentice-io and archimetis each named no provider at all and took None. Beam names providers, publishes a directory and exposes the choice in the builder.

WHY CONFIDENCE IS 0.7 RATHER THAN HIGHER, STATED PLAINLY: the core-concepts page names the model parameter without enumerating the selectable models, the SELECTING TOOLS and per-provider LLM pages were not opened within budget, and the two providers named appear in an FAQ answer rather than a product specification. The SELECTION SURFACE is documented; the SELECTABLE SET is not. That gap bears on confidence and not on the grade, since the axis asks who chooses. CONFIDENCE 0.7.

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

CORRECTED Partial to Full, AND THE APRIL PARTIAL WAS A RETRIEVAL FAILURE RATHER THAN A JUDGEMENT — the cell cited a rubric, not a search, and the surface it missed is one of the deepest in this lane. THE FULL BAR IS A DOCUMENTED API OR SDK FOR THE VENDOR'S OWN PLATFORM AND IT IS MET SEVERAL TIMES OVER.

A public REST API is documented at docs.beam.ai/08-reference/api with a stated base URL of api.beamstudio.ai, API-key authentication via an x-api-key header plus a current-workspace-id header, a dedicated Authentication page, and worked code examples. AN OPENAPI SPECIFICATION IS PUBLISHED at /openapi-final.json, which is the machine-readable form of the same claim. THE ENDPOINT SET IS NOT A READ-ONLY VENEER — IT BUILDS AGENTS PROGRAMMATICALLY.

CREATE A COMPLETE AGENT GRAPH creates a new agent with its full graph including nodes, edges and tool configurations, producing 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 execution flow; UPDATE PROMPT OF THE TOOL and UPDATE INPUT/OUTPUT PARAMETERS reach inside a node's configuration; PUBLISH AGENT GRAPH ships it. A caller outside the product can author, wire, configure, test and publish an agent without opening the UI.

That is the platform being callable, which is the axis. RUNTIME IS EXPOSED TOO: 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. AN MCP CONNECTION IS DOCUMENTED as its own reference page, and a skill can be imported into an agent from a zip file.

THE VENDOR STATES THE POSTURE DIRECTLY: THE PLATFORM IS MODEL-AGNOSTIC AND API-FIRST, SO YOU CAN EVOLVE YOUR AGENTIC WORKFLOWS WITHOUT LOCK-IN, and elsewhere EMBED AGENTS VIA COMPONENTS AND APIS, MANAGE TENANTS AND USAGE, AND ENFORCE POLICIES CENTRALLY. THE DIRECTION TEST THIS AXIS TURNS ON IS SATISFIED IN THE RIGHT DIRECTION: these endpoints let an external system COMMAND Beam, not merely notify it — the distinction that held archimetis, rogo, sana and shortwave at Partial or None. A CUSTOM INTEGRATIONS surface additionally lets a customer build an integration for an internal or proprietary platform. CONFIDENCE 0.85.

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

Full CONFIRMED and re-based off the rubric, AND IT CLEARS BOTH HALVES OF THE BAR: a change is genuinely under test, and the result is readable. THE PRE-DEPLOYMENT HARNESS IS EXPLICIT AND USES REAL DATA: TEST MODE lets a builder RUN AN AGENT FLOW WITH DATA FROM PAST TASKS IN A SAFE ENVIRONMENT BEFORE YOU PUBLISH.

Past production tasks replayed against an unpublished draft, in isolation, before release, is the definition of a change under test — and it pairs with the draft-versus-published graph model to make the gate structural rather than advisory. BACKTESTING IS NAMED AS SUCH: RERUNNING TASKS covers RE-EXECUTE TASKS FOR DEBUGGING, TESTING WORKFLOW CHANGES, BACKTESTING PROMPT IMPROVEMENTS, AND VALIDATING AGENT PERFORMANCE. Backtesting a prompt improvement is comparing a modified agent against known cases.

THE RESULT IS A READABLE SCORE, WHICH IS WHAT SEPARATES FULL FROM A QUALITY GATE WITH NO OUTPUT: the EVALUATION FRAMEWORK page covers VALIDATE AGENT ACCURACY WITH EVALUATION CRITERIA, AUTO-RUN RETRIES, AND ANALYTICS MONITORING, and the Analytics dashboard surfaces EVALUATION SCORES beside completion rates. Named criteria producing a score a human reads is the comparable result the bar asks for.

EVALUATION IS BOUND TO THE STEP, NOT BOLTED ON: the Flow node sidebar has three tabs — PARAMETERS, SETTINGS, AND EVALUATION — so criteria are attached to individual tool nodes, and Auto-run retries means a failed criterion changes runtime behaviour rather than only reporting.

A CLOSED IMPROVEMENT LOOP IS DOCUMENTED AND PROGRAMMATIC: RATE TASK OUTPUT submits A RATING FOR TASK EXECUTION OUTPUT TO IMPROVE AGENT PERFORMANCE; OPTIMIZE TOOL will OPTIMIZE A TOOL'S PERFORMANCE AND CONFIGURATION BASED ON USAGE PATTERNS AND FEEDBACK, with GET TOOL OPTIMIZATION STATUS to track it; and an OPTIMIZE OUTPUTS page documents PRO LEARNING TOOLS TO TEST, TUNE, AND OPTIMIZE AN AGENT'S TOOL OUTPUTS.

The index's own editorial records the same shipping as a LEARNING HUB that aggregates agent feedback and self-checks, spots recurring failures and DRAFTS THE INSTRUCTION FIX FOR REVIEW — a controlled post-deployment optimisation loop with a human gate on the change. NOT COUNTED: the score being visible on a dashboard is Obs; the human approving a drafted fix is HITL. CONFIDENCE 0.85.

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

None CONFIRMED and re-based off the rubric, and the April cell reached the right answer at 0.6 with no source. NONE OF THE THREE MODALITIES IS DOCUMENTED — no hosted or local browser session the agent drives, no desktop control, no remote or local computer control — across a documentation estate deep enough that the absence is positive rather than inferred. The docs enumerate the step types a Flow can contain: AI, INTEGRATION, CONTROLS, OR UTILITIES.

There is no browser step, no screen step and no computer-use step among them, and no page in the 01 to 08 documentation tree addresses operating an interface. THE ARCHITECTURE ANSWERS THE QUESTION DEFINITIVELY BECAUSE THE ALTERNATIVE IS THE WHOLE PRODUCT: Beam reaches other systems through connectors and APIs.

INTEGRATIONS, VIA INTEGRATION CONNECTORS, ARE THE TOOLS THE AGENT USES INSIDE WORKFLOW STEPS TO TAKE ACTION IN EXTERNAL SYSTEMS, FOR EXAMPLE SENDING AN EMAIL, CREATING A RECORD, OR FETCHING DATA FROM APPS LIKE GMAIL, SALESFORCE, AIRTABLE, OR SLACK, and the platform is described as API-FIRST. THE BUYER QUESTION THE AXIS NAMES HAS NO ANSWER HERE: how often does it break when UI elements change is meaningless against a Salesforce API connector.

Salesforce could redesign its entire interface and a Beam Flow would be unaffected. THE NEAR MISS IS SUBSTANTIAL AND IS REFUSED FOR THE RIGHT REASON, WHICH IS WHY IT IS RECORDED. Beam positions itself as the successor to robotic process automation and an FAQ lists RPA AGENTS: CAPTURE INVOICES, POST ORDERS, RECONCILE PAYMENTS, AND VALIDATE FORMS among AI agent examples.

That passage is a taxonomy of the general category — what AI agents do across the market — not a claim about a Beam capability, and nothing on any product page or documentation page says a Beam agent drives a screen. Claiming the RPA lineage is positioning; operating a GUI is a capability, and only one of them is documented.

THE INDEX'S OWN COMPARISON PROSE MAKES THE SAME DISTINCTION INDEPENDENTLY, describing a rival platform's COMPUTER USE CAPABILITY THAT LETS AGENTS RELIABLY OPERATE GUIS FOR APPS WITHOUT APIS as the thing that stands out about it, in an article published on Beam's own blog. CONFIDENCE 0.8.

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