EverWorker vs Joget (2026)

Stands at F, re-based onto the current product and strengthened by the mechanism the July basis did not name. THE UNIVERSAL CONNECTOR IS THE LOAD-BEARING FACT: upload an OpenAPI spec or connect via REST or GraphQL, and the platform AUTO-DISCOVERS PERMITTED ACTIONS in the customer's systems. Generating an agent's action surface from a machine-readable contract is a different and better trade than a hand-built catalogue: coverage is bounded by what the customer's systems publish rather than by what the vendor got round to building, and it stays current with the spec. SCOPING BY PERMISSION IS THE PART WORTH CARRYING, and it is unusual. The connector discovers PERMITTED actions, not all actions, so the tool surface is bounded by the credentials in use. Most platforms in this lane expose everything a connector can reach and leave authorisation to the target system at call time. NAMED INTEGRATIONS COVER GENUINELY DIFFERENT CLASSES: model providers including OpenAI, Anthropic, OLLAMA, Vertex and Fireworks; productivity in Google and Microsoft; messaging in Telegram; project and web tooling in ClickUp and WIX; and knowledge sources across Google Drive, OneDrive, Slack and CRM. THE THIRD ROUTE IS THE ONE CREDITED ON COMP AND IS RELEVANT HERE ONLY AS SCOPE: where a system has no API at all, the AI agent browser reaches it. Between a spec-driven connector and a browser, there is little the platform cannot touch, which is what makes the breadth claim credible rather than aspirational. What is not documented is an enumerated connector directory or count, so class breadth rests on the named examples plus the generative connector rather than on a catalogue.

Stands at F, with the tool palette now enumerated rather than described as a plugin architecture. THE AGENT TOOL SET IS NAMED: DATABASE QUERY TOOL, JSON API TOOL, EMAIL TOOL, GOOGLE WEB SEARCH TOOL, BEANSHELL TOOL, RUN AI AGENT TOOL, MCP CLIENT TOOL and PINECONE TOOL. Those span genuinely different classes: the customer's own database, arbitrary HTTP APIs, messaging, the open web, custom code, other agents, MCP servers and a vector store. THE BEANSHELL TOOL IS THE CEILING-REMOVER and is unusual in a no-code builder. It executes custom Java-syntax code as an agent tool, so anything the JVM can reach is reachable without waiting for a connector. Combined with the open-source core and plugin architecture, integration coverage is bounded by what the customer will build rather than by the vendor's roadmap. THE MCP CLIENT TOOL IS CREDITED HERE AND NOT ON EXT, per the standing one-line test: Joget calling remote MCP servers is the vendor's product consuming third-party tools. The same correction was applied to wassist and agentx today. It matters more here than usual, because it means any MCP-exposed service becomes agent-callable without Joget building anything. UNDERNEATH SITS THE PLATFORM'S OWN INTEGRATION LAYER, an Application and Integration Fabric connecting legacy and modern systems through prebuilt connectors and an extensible plugin architecture, which predates the agent work by many years and is the reason a Joget agent inherits reach rather than starting from zero. What is not documented on the pages reached is an enumerated connector catalogue or count for the platform layer, so class breadth at that level rests on the fabric description while the agent tool set is enumerated precisely.

Stands at F, re-based off trade press and onto the vendor's own material, with the current product's architecture named. THE TWO-CLASS MODEL IS THE ORCHESTRATION PRIMITIVE and it is a genuine design rather than a marketing split. UNIVERSAL WORKERS are digital knowledge workers available out of the box that augment a team; SPECIALIZED WORKERS are tightly defined agents that FOLLOW STRICT WORKFLOWS AND GUIDELINES WHICH THE USER DEFINES IN A FLOWCHART-LIKE INTERFACE. One class is open-ended and reasons; the other is bounded and repeats. Composing a process from both is how this platform expresses orchestration. THE FLOWCHART INTERFACE IS THE CONTROL-FLOW SURFACE, which most Full cells in this lane lack. A user-defined flowchart is inspectable and constrains execution order, so a customer can see how a process runs rather than trusting the model to decide each time. MULTI-AGENT COMPOSITION IS DOCUMENTED at the layer above: agents are grouped into applications sharing a COMMON CONTEXT AND GOAL, described as hierarchical pipelines. Shared context across a group is what allows a handoff to carry meaning rather than restating the problem, and it is the same structure credited as working memory on Mem. THE FRAMING IS WORKFORCE RATHER THAN WORKFLOW and that is consistent throughout: departments configure department-specific AI workforces, and the vendor's positioning is always-on work rather than triggered automation. That shapes what the orchestration is for, and it is why the trigger coverage sits lower than the orchestration. What is not documented is branching, looping or conditional vocabulary by name, nor failure and retry behaviour when a worker errors mid-process. The flowchart interface implies control flow; no page reached enumerates its constructs.

Stands at F, and this record has two orchestration layers rather than one, which is what makes it strong. THE AGENT LAYER IS A DOCUMENTED THREE-PART TASK MODEL. An agent has a brain, where the LLM service and default prompts are set, and one or more TASKS THAT RUN IN SEQUENCE, each composed of a PROMPT, an optional TOOL that calls an API, retrieves data or runs a database query, and an optional ENHANCER that stores or transforms the output. Multiple tasks execute in order, so an agent is a pipeline rather than a single call, and the enhancer step means each task's output is captured for the next. THE PROCESS LAYER IS THE OLDER AND DEEPER ONE. Joget's lineage is a workflow and process automation engine, and the AI AGENT TOOL embeds an agent as a step inside a business process. So a composed system is agents inside processes: deterministic routing, human tasks and agent tasks in one flow, with process variables carrying state between them. THAT SPLIT IS THE ARCHITECTURAL POINT WORTH CARRYING. Where most builders in this lane implement orchestration for agents from scratch, Joget attaches agents to an existing BPM engine. The reasoning layer is new; the sequencing, state, participation and audit machinery is twenty years old. For processes that must run the same way every time with an agent at one step, that is a different and often better shape than a purely agentic graph. A RUN AI AGENT TOOL lets one agent invoke another, giving multi-agent composition inside the agent layer as well. What is not enumerated on the AI pages reached is branching or conditional vocabulary within an agent's own task sequence; control flow of that kind lives in the process engine rather than in the agent.

Stands at P, re-based, with the honest position that this is the axis the product's own framing works against. WHAT IS DOCUMENTED. Integrations that carry inbound work exist in Telegram, Gmail and CRM systems, so messaging and email are real entry routes. The vendor's central claim is an ALWAYS-ON agentic workforce, which implies continuous rather than invoked operation, and Universal Workers ASSIST EMPLOYEES, which implies a person addressing a worker directly. WHY IT HOLDS AT PARTIAL AND NOT HIGHER. Always-on is a description of availability, not a trigger mechanism, and I will not credit a class from an implication. No scheduling, cron or recurrence capability is named. No inbound webhook or event subscription appears as a distinct capability. No enumerated channel surface exists: there is no documented chat, web widget, Slack or embedded interface through which a person reaches a worker, beyond the integrations above and the platform's own UI. THE UNIVERSAL CONNECTOR CUTS BOTH WAYS and is worth naming. Auto-discovering permitted actions from an OpenAPI spec builds the outbound action surface; it says nothing about how work arrives. Several records this session had that reversed, crediting an integration catalogue as trigger coverage, and this basis should not repeat it. CONFIDENCE MEDIUM AND THE EXPECTATION RECORDED: a platform selling always-on finance and HR workers that reconcile data and run invoicing approvals is almost certainly scheduling something, and a triggers or automation page would likely move this cell. Nothing of the kind was reached across two passes, and the vendor's product surface is thin on mechanism generally, which is the pattern across this whole record.

Stands at P, and the shape here is that the platform is rich where the agent documentation is quiet. WHAT IS DOCUMENTED FOR AGENTS. The AI AGENT TOOL makes an agent a step in a business process, so anything that starts a process starts an agent. Joget's process engine has long carried scheduled and event-driven process initiation, form submission triggers and API-triggered processes, and agents inherit all of it by being process steps. An EMAIL TOOL gives an outbound channel, and agents are embedded into applications so the app's own user interface is an inbound surface. WHY IT HOLDS AT PARTIAL RATHER THAN RISING. Inheriting the platform's triggers is a strong argument and I have not credited it as if it were documented for agents. No page reached enumerates trigger types for an AI agent specifically: no scheduling or cron configuration, no inbound webhook, no event subscription and no messaging or chat channel appears in the AI Agent Builder documentation. Crediting a class because the surrounding platform probably supports it would be the inference-over-evidence error corrected elsewhere in this batch. THE ASYMMETRY IS CHARACTERISTIC OF THIS RECORD and worth carrying. Joget's agent documentation is unusually precise about components, prompts, tools, enhancers, LLM services, audit trails, and unusually thin about how work arrives. That is what you would expect from a mature process platform where the trigger question was answered years ago and nobody thought to restate it for agents. CONFIDENCE MEDIUM AND THE EXPECTATION RECORDED: the process documentation almost certainly enumerates the trigger types, and reading it against the AI Agent Tool is the cheapest check on this record. My expectation is that this moves to Full.

Stands at F and is one of the better-evidenced Know cells in the lane, because the vendor's differentiation claim is precisely on this axis. THE ENTERPRISE KNOWLEDGE ENGINE IS A MAINTAINED STRUCTURE, not per-request assembly. The vendor states there is NO NEED TO CREATE A DATA WAREHOUSE OR RAG INFRASTRUCTURE: the platform connects directly to business systems and stores relevant information in ITS PROPRIETARY ENTERPRISE KNOWLEDGE GRAPH AND CONTEXT ENGINES, WITH NO ETL OR DATA ENGINEERING REQUIRED. A knowledge graph is the strongest form this axis recognises, since it encodes relationships between entities rather than similarity between chunks. THE ADOPTION MECHANIC IS THE DETAIL WORTH CARRYING and it is well judged for the buyer. Assigning knowledge to a new worker is a toggle: JUST TOGGLE ON COMPANY DOCS AND DATA WITH A CLICK. The vendor's framing is to HAND YOUR AI WORKERS THE SAME GUIDES, PLAYBOOKS, AND DOCUMENTATION YOU'D GIVE A NEW HUMAN HIRE, which is the right mental model for a non-technical buyer and removes the step where most no-code grounding projects stall. PERMISSION-AWARE RETRIEVAL IS THE PROPERTY THAT MATTERS MOST FOR AN ENTERPRISE and few records in this lane document it: knowledge is gathered RESPECTING EXISTING PERMISSIONS, so a worker grounded on company systems cannot surface material the requesting user could not otherwise see. That control is credited on Sec; noted here because it is what makes broad ingestion safe rather than reckless. Sources sync from Google Drive, OneDrive, Slack and CRM, and Smart Search agents run semantic search over the result. What is not documented is embedding model, chunking, re-index cadence, or citation of sources in answers.

P>F. The July basis said no explicit first-class RAG store was documented. The component reference names one. A PINECONE TOOL IS IN THE DOCUMENTED TOOL PALETTE, alongside the Database Query Tool and Google Web Search Tool. Pinecone is a vector database, so a first-class Pinecone tool is a documented retrieval integration against a persistent index the customer maintains. Under the persistence line that is a maintained retrieval structure rather than per-request assembly, and under the 31 August ruling a capability the customer wires from documented primitives counts. THE PROMPT PALETTE IS THE SECOND HALF AND IS MORE INTERESTING THAN THE VECTOR STORE. Grounding sources are enumerated as components: FILE PROMPT for documents, FORM DATA PROMPT and LIST DATA PROMPT for the application's own records, FORM DATA TABLES SCHEMA PROMPT for the data model itself, and JSON API PROMPT for live external data. Feeding an agent the schema of the customer's tables, not just their contents, is unusual and is what lets an agent reason about where information lives rather than only about what it was handed. THE STRUCTURED HALF IS WHERE THIS PLATFORM IS GENUINELY STRONG. Most records in this lane ground on documents; Joget grounds on the application's own forms, lists, tables and database through a Database Query Tool, which for a platform whose customers build the applications is grounding on the system of record itself. What is not documented on the pages reached is chunking, embedding model, index maintenance or citation of sources in answers, and Pinecone is a customer-supplied dependency rather than a bundled store. Those are the limits; they do not change that a persistent retrieval structure is a documented, first-class part of the builder.

Stands at P under the state limb of the 31 August ruling, and the vendor's own word choice is what holds it there. THE PLATFORM CALLS THEM MEMORIES AND THEN DEFINES THEM AS SOMETHING ELSE: assigning knowledge to a worker is a toggle, and THESE MEMORIES SERVE AS WORKING CONTEXT TO DRIVE CONSISTENT, RELEVANT OUTPUTS. Working context is what is supplied to a worker at run time from a maintained corpus; it is grounding wearing memory's name. That corpus is the Enterprise Knowledge Engine and is credited on Know, and counting it here would be the double-count this record otherwise avoids. WHAT DOES SUPPORT PARTIAL IS SHARED STATE ACROSS A GROUP. Agents are grouped into applications that share A COMMON CONTEXT AND GOAL, so state is held across the members of a running application rather than isolated per agent. That is real state, and it is more than a single conversation buffer, which is why this sits at Partial rather than lower. WHY IT DOES NOT REACH FULL. Nothing documents state that persists across sessions and is addressable by the agent: no memory store, no retention policy, no per-user or per-customer memory keyed across runs, and no primitive by which a worker writes something in one execution and reads it in the next. The ruling requires exactly that, by either the accumulation or the state route, and neither is documented. THE TENSION WORTH RECORDING FOR THE PRODUCT. EverWorker's whole framing is an always-on workforce whose members work like a colleague from day one. A colleague who remembers nothing between tasks is a weaker proposition than the framing implies, so this is likely undocumented rather than unbuilt, and it is the cell most worth checking at lane close alongside the developer surface.

P>F under the 31 August ruling, which allows Full where the customer wires persistence from documented primitives rather than requiring automatic accumulation. THE PRIMITIVE IS NAMED AND IT IS A MEMORY PRODUCT: the AI Agent Builder's tool palette includes a MEM0 CHAT MEMORY TOOL. Mem0 is a purpose-built persistent memory layer for agents, not a generic store, so a vendor listing it as a first-class tool is shipping cross-session memory as a documented component rather than leaving the customer to invent one. The July basis said no distinct conversational agent memory was documented; this is exactly that. TWO ENHANCERS PROVIDE THE SECOND, MORE JOGET-NATIVE ROUTE. STORE TO FORM ENHANCER and STORE TO PROCESS VARIABLE ENHANCER write agent output into the application's own data layer and into process variables. That is durable state in the customer's database, addressable by later tasks, later process steps and other agents, and it persists for the life of the record rather than the conversation. THE ARCHITECTURE IS WHY THIS IS STRONGER THAN IT LOOKS. Because agents are embedded in Joget processes, an agent's output becomes application data that a subsequent human task, a later agent run or a scheduled process can read. State outliving the run is the default rather than a feature, which is the advantage a BPM lineage gives here. A RUN AI AGENT TOOL lets one agent invoke another, so state passes between agents within a composed system. WHAT IS NOT DOCUMENTED on the pages reached is retention configuration for Mem0-held memory, or a memory inspection and deletion surface. Confidence stays high because the primitives are enumerated in the vendor's own component list rather than described in prose.

N>P, and the July basis made a specific error worth naming: it graded from positioning. Its reasoning was that the platform IS POSITIONED AS AUTONOMOUS WITH LITTLE OR NO HUMAN OVERSIGHT, sourced partly to trade press. How a vendor markets autonomy is not evidence about what constraint mechanisms it ships, and inferring absence from marketing tone is the mirror image of crediting capability from marketing prose. WHAT IS ACTUALLY DOCUMENTED IS A CONSTRAINT SURFACE. Agents come in two declared forms, and the constrained one is explicit: SPECIALIZED WORKERS FOLLOW STRICT WORKFLOWS AND GUIDELINES WHICH THE USER DEFINES IN A FLOWCHART-LIKE INTERFACE. A user-defined flowchart that an agent must follow is a real bound on behaviour, and offering it as a distinct worker class means the autonomy level is a design-time choice rather than a fixed property. THE SECOND CONSTRAINT IS THE PERMISSION SCOPE credited on Sec: agents operate within existing system permissions and the connector exposes only permitted actions, so what an agent can do is bounded by what its credentials allow. WHY IT STOPS AT PARTIAL. Nothing pauses a worker mid-run pending a person's approval. No checkpoint, pending-action queue, confidence threshold or reviewer surface appears in any product material reached across two passes. For a platform whose stated purpose is always-on autonomous work inside finance and HR systems, an approval step before a consequential write is the mechanism a buyer would ask for, and it is not documented. A SOURCING TRAP RECORDED: the vendor's blog discusses keeping humans in approvals for bank changes and garnishments, and building approval tiers into playbooks. That is buyer guidance about how to run AI safely, not documentation of a product feature, and it is not credited. The same blog series produced the SOC 2 trap noted on Sec.

Stands at F, and the mechanism is structural rather than a feature the vendor added for agents. THE GATE IS THE PROCESS ENGINE ITSELF. Joget's lineage is workflow and business process management, where a human activity assigned to a participant is the primitive the whole product is built on. The AI AGENT TOOL places an agent as a step inside such a process, so an agent task and a human approval task are the same kind of object in the same flow. Routing work to a person for a decision is not an oversight feature bolted onto an agent runtime; it is what the runtime was always for. THE VENDOR MAKES THE CLAIM DIRECTLY, describing agents as automating processes within structured, auditable workflows that keep humans in control. On most records I would treat that as governance prose and refuse it, and I would be right to. Here it is redeemed by the mechanism sitting underneath it: process participation is a documented, audited concept, to the point that DX 8's security audit checks PROCESS PARTICIPATION as one of three things it inspects. THE GOVERNANCE APP CORROBORATES from the other direction. Its administration documentation notes that privileged access BYPASSES THE NEED FOR APPROVALS, which is only a meaningful statement if approvals are a normal part of the flow. A SECOND, WEAKER GUARDRAIL LAYER exists in the agent itself: default prompts and a fixed task sequence constrain behaviour, and enhancers determine what is persisted. CONFIDENCE IS MEDIUM RATHER THAN HIGH for one reason: I did not reach a page documenting an approval step placed specifically around an agent task, as opposed to the process engine's general human activity. The inference is strong given the AI Agent Tool is a process step, but it is an inference about composition rather than a documented pattern.

N>P. The July basis recorded that no certifications, SSO, RBAC or audit tooling were documented IN RETRIEVED SOURCES, which was honest about its own limit, and None asserts more than that. A named control does exist. THE CONTROL IS PERMISSION INHERITANCE, which section 7 names explicitly among the things an F bar looks for. The vendor states EverWorker CONNECTS DIRECTLY TO ALL BUSINESS SYSTEMS, RESPECTING EXISTING PERMISSIONS, and separately that the Universal Connector AUTO-DISCOVERS PERMITTED ACTIONS in the customer's systems. Inheriting the source system's permissions rather than operating on a service account is the right architecture and the strongest single control for this product shape, because it means the platform cannot become a route around a customer's existing access model. AUTO-DISCOVERING ONLY PERMITTED ACTIONS is the second half of the same idea applied to tools: the agent's action surface is bounded by what its credentials already allow, rather than by what the connector could theoretically call. WHY PARTIAL AND NOT HIGHER. The attestation half is entirely absent. No SOC 2, ISO, trust page, audit report, SSO or retention configuration was reached across two passes. Under the conjunction bar, one half documented and the other absent is Partial. A TRAP RECORDED SO IT IS NOT MISREAD LATER. EverWorker's blog carries buyer-guidance posts advising readers to REQUIRE SOC 2 TYPE II OR ISO/IEC 27001, documented data flows, subprocessor transparency and DPAs from vendors. That is advice to buyers about what to demand, not a claim about EverWorker's own posture, and reading it as the latter would credit the vendor with certifications it has not asserted. Vendor blogs that publish procurement checklists are an easy place to make that mistake. Deployment location is graded on Dep and not borrowed here.

Stands at F, and both halves of the conjunction are now documented rather than one. THE ATTESTATION IS STATED AS CERTIFICATION, NOT ALIGNMENT. The vendor's own site footer reads ISO/IEC 27001:2022 CERTIFIED. That is a claim of a certificate issued against the current version of the standard, and it is the top of the hedge ladder rather than a rung on it. THE CONTRAST WITH ALTILIA, REVIEWED IMMEDIATELY BEFORE THIS RECORD, IS WORTH KEEPING because the two illustrate the boundary cleanly. Altilia states a management system ALIGNED WITH ISO/IEC 27001 and sits at Partial; Joget states ISO/IEC 27001:2022 CERTIFIED and sits at Full. Same standard, one word apart, and the word is the whole difference. Neither vendor was read generously or harshly; each was read literally. THE CONTROL HALF IS DEEP AND PRODUCT-SPECIFIC. Role-based access control and audit trails are core platform properties; DX 9 adds built-in monitoring, audit trails and a Governance Center as a centralised portal for managing and monitoring installations, apps and developers; and DX 8 introduced one-click security, performance and governance audits covering user authentication, user interface access and process participation. THE GOVERNANCE AUDITS ARE THE UNUSUAL PART and worth carrying. Most platforms in this lane document controls the customer configures. Joget ships a check that inspects the customer's own applications for authentication and access weaknesses, which is security tooling aimed at what citizen developers build rather than only at the platform hosting them. For a low-code platform, where the risk is a business user publishing something over-permissive, that is the right place to point it. One fact deliberately excluded: a partner site claims FISMA, PCI DSS and DISA compliance. That is a third-party integrator's page, not the vendor's, and is not credited. Deployment breadth is graded on Dep.

Stands at P but confidence drops from medium to low, because re-reading the evidence shows less than the July basis implied. WHAT THE BASIS ACTUALLY RESTED ON was that AI workers EXECUTE REAL WORK AND REPORT THEIR PERFORMANCE. Reporting performance is an outcome claim, not an observability surface: it says a worker will tell you how it did, not that a customer can reconstruct what it did. The vendor's framing of workers as teammates makes that language natural and it should not be read as tooling. WHAT I FOUND THIS PASS DOES NOT MUCH IMPROVE IT. A blog post states that customers deploy AI workers running end-to-end workflows WITH GOVERNANCE, OBSERVABILITY, AND CONTINUOUS LEARNING BUILT IN. Observability appears as a noun in a list, with no mechanism attached, and it comes from marketing-adjacent blog content rather than product documentation. Under section 7 that is thought-leadership prose, and I am not crediting a capability from a word in a sentence. WHAT IS ABSENT IS EVERYTHING THE AXIS ASKS FOR. No trace, run history, execution log, action record, reasoning trail, audit log, dashboard, retention period or export path appears on any surface reached across two passes. WHY IT REMAINS PARTIAL RATHER THAN NONE. Performance reporting is asserted directly by the vendor as something workers do, and it is at least a customer-facing output about agent behaviour. That is thin, and the low confidence now records how thin. THE GAP IS MOST CONSEQUENTIAL ON THIS RECORD OF ANY THIS SESSION. EverWorker sells always-on autonomous workers that reconcile finance data, run invoicing approvals and act in HR systems, and this pass credited it with a browser that operates websites directly. An agent with that reach and no documented audit trail is the combination a buyer should ask about first, and it is the second open item at lane close after the developer surface.

P>F. The July basis rested on governance health checks and a Governance Center and noted that per-agent run tracing was less detailed. The documentation has a page for exactly that, titled AGENT EXECUTION AUDIT TRAIL. TWO NAMED DOCUMENTATION PAGES CARRY THIS CELL: AGENT EXECUTION AUDIT TRAIL and SYSTEM AGENTS AND TRACEABILITY, both sitting in the AI Agent Builder section of the knowledge base alongside the component reference. A vendor writing a dedicated page on agent execution audit trails has shipped the artefact this axis asks for, and the naming is unambiguous in a way that governance and monitoring language is not. THE PLATFORM LAYER UNDERNEATH IS UNUSUALLY DEEP for this lane and it is the reason an agent audit trail is credible here rather than aspirational. Joget is a process automation platform with audit trails and RBAC as core properties, DX 9 adds built-in monitoring, audit trails and a Governance Center, and DX 8 introduced one-click security, performance and governance audits covering user authentication, user interface access and process participation, plus checks for unusually high error rates and data integrity in running processes. THE ARCHITECTURAL POINT WORTH CARRYING. Because agents execute as steps inside governed processes rather than as a separate runtime, an agent's actions land in the same process history as human tasks. An auditor reviewing a case sees the human approvals and the agent executions in one trail, which is a materially better answer than a separate agent dashboard bolted beside a workflow log. WHAT I DID NOT REACH is the audit trail page's contents, so what a record contains, its retention and whether it can be exported rest on the page's title and the platform's general audit capability rather than on read detail. That is the check at lane close.

Stands at F. Both modes are documented consistently, including on a third-party review surface where the vendor's own description is reproduced. SAAS AND CUSTOMER-CONTROLLED DEPLOYMENT ARE BOTH STATED: deploy EverWorkers WITHIN YOUR CHOSEN INFRASTRUCTURE, ON-PREMISE OR PRIVATE CLOUD, MAINTAINING FULL CONTROL OVER DATA PRIVACY AND SECURITY, alongside the hosted option. Customer-chosen infrastructure is the top of this axis, since both where the software runs and where the data sits become the customer's. THE MODEL SIDE MAKES IT COHERENT RATHER THAN NOMINAL, which is the test that separated the strong from the weak Dep cells this session. Because the platform supports OLLAMA and locally served open-weight models, an on-premises customer is not obliged to ship prompts to a hosted provider for inference. A deployment claim whose inference still leaves the perimeter answers only half the question, and this one does not have that hole. THE KNOWLEDGE ARCHITECTURE POINTS THE SAME WAY. The Enterprise Knowledge Engine connects directly to business systems rather than requiring an ETL pipeline into a vendor warehouse, so the footprint is smaller by design. CONFIDENCE IS MEDIUM AND THE LIMIT IS SPECIFIC. No installation guide, infrastructure requirement, statement of which components run customer-side, or region selection for the SaaS offering was reached across two passes. Under the test settled on ai-library today, a capability can exist without being offered, and what is missing here is the route rather than the claim. It stays at Full because the claim is repeated across the launch announcement, the product page and an independent listing, and because the enterprise motion is services-led, which is a plausible route in itself. Deployment is graded here alone; the permission model is credited on Sec.

Stands at F and is among the strongest Dep cells in the lane, because the open-source core makes self-hosting a right rather than a tier. THE SPREAD IS COMPLETE: the vendor states the platform DEPLOYS ANYWHERE, CONTAINERIZED, CLOUD-NATIVE, OR ON-PREM, and is infrastructure agnostic across private cloud, public cloud, hybrid and on-premises. Certified cloud deployments on AWS, Azure, Red Hat and Google Cloud are documented. THE OPEN-SOURCE COMMUNITY EDITION IS THE PART THAT SEPARATES THIS FROM A CLAIM. Most on-premises options in this lane are enterprise-tier privileges obtained through a sales conversation, and today I held two records at Partial for exactly that, a capability asserted with no route for a customer to take it. Here the route is downloading the source. A customer can run the platform in their own data centre without contacting the vendor, which is the difference between an option and an entitlement. THE MODEL SIDE CLOSES THE RESIDENCY LOOP, which is the test that separated the strong from the weak Dep cells this session. Because agents use a customer-supplied LLM key and a proxy domain can route model calls through the customer's own gateway, an on-premises deployment is not obliged to send prompts out through a vendor-managed inference path. The customer chooses both where the platform runs and where the model calls go. ONE COMMERCIAL LIMIT WORTH RECORDING and it is real: the AI Agent Builder is an Enterprise Edition feature in DX 8.1 and later, and comes by default in DX 9.1 and above. So the platform self-hosts freely; the agent builder specifically requires a paid edition. The deployment freedom is genuine, the agent capability is licensed. What is not documented on the pages reached is region selection within the vendor's own cloud offering.

Stands at F under the 31 August bar, which asks whether the customer receives packaged assets ready to adopt, and it clears it by two distinct routes. THE FIRST IS THE UNIVERSAL WORKER, which is the strongest form on this axis: these workers are AVAILABLE OUT-OF-THE-BOX, WITH THE ABILITY TO DEEPLY UNDERSTAND CORPORATE KNOWLEDGE AND ASSIST EMPLOYEES FROM DAY ONE. A worker that is useful on day one without configuration is a finished asset, not a scaffold, and the knowledge toggle is what makes that credible: the packaging is generic and the grounding is the customer's. THE SECOND IS DEPARTMENTAL: EXPERT-DESIGNED, KPI-FOCUSED AI WORKERS, SET UP BY OUR TEAM, TUNED FOR YOURS, with department-specific workforces for finance, operations and HR ready to configure to the customer's data and tools. Named finance examples include fraud reduction, data reconciliation, audit readiness, and invoicing, approvals and payments across vendors and customers. A LIBRARY OF MORE THAN TWENTY PREBUILT AGENTS sits underneath as the component layer. ONE HONEST TENSION RECORDED. Expert-designed and set up by our team edges toward the engagement rather than the platform, and the 31 August ruling excludes capability supplied by the vendor's people. It does not sink this cell, because Universal Workers are explicitly out-of-the-box and the prebuilt library exists independently, but a buyer should know that the departmental workforces arrive with services attached rather than as self-serve packs. What I did not reach is an enumerated catalogue, so the twenty-plus figure and the departmental set rest on the vendor's description rather than on a browsable surface.

Stands at P, and the July basis put it correctly: the marketplace is real and agent-specific packs are what is missing. WHAT EXISTS IS A GENUINE BROWSABLE CATALOGUE. The Joget Marketplace is an app store for enterprise apps, plugins and templates, and it is not a claim: it is where the AI Agent Builder plugin itself is downloaded from, which the installation documentation walks through step by step. A catalogue the vendor ships its own product through is unambiguously operational. WHY THAT DOES NOT CARRY FULL. The bar asks what the customer receives as packaged assets for the axis in question. Apps, plugins and templates are platform artefacts; a plugin is a capability the customer installs, and an app template is a starting application. Neither is a prebuilt agent. No library of ready-made agents, no agent template gallery and no packaged departmental agents appear on any page reached across two passes, which is what would move this cell. WHAT SITS CLOSER THAN THE MARKETPLACE is inside the builder: the component palette supplies reusable named building blocks, including a PERSONA PROMPT for defining an agent's role and a documented BEST PRACTICES FOR PROMPTING AND COMING UP WITH AGENTS page. Those accelerate building an agent; they are not an agent someone else already built. THE TENSION WORTH RECORDING: this is a platform with an established third-party plugin ecosystem and a first-party agent builder that shipped recently. Agent packs are the obvious thing for that marketplace to accumulate, and the July note's phrase, agent specific packs are emerging, is likely to age into a correction. Checking the marketplace for an agent category is the cheap test at lane close, alongside the process trigger types.

Stands at F, re-based off press coverage, though it is the least directly documented of the Full cells and the note should say so. WHAT SUPPORTS IT. The vendor states support for all major LLMs, with named providers spanning OpenAI, Anthropic, Google Vertex and Fireworks, and OLLAMA. OLLAMA is the entry worth noting: it runs open-weight models locally, so the platform's model support extends past hosted APIs to models a customer hosts themselves. Combined with the on-premises deployment credited on Dep, that gives a buyer a path where neither the platform nor the model leaves their infrastructure. THE SELECTION MECHANIC IS DOCUMENTED AND UNUSUAL. Built-in benchmarking tests agents against base models SO TEAMS CAN PICK THE MOST COST EFFECTIVE AND ACCURATE LLM FOR EACH JOB, and per each job means selection is per agent rather than per account. A platform that helps a customer decide which model to use, with cost and accuracy side by side, is doing more than exposing a dropdown, and it is the same feature graded down on Eval for measuring the model rather than the agent. That is the correct place for it to earn credit. CONFIDENCE IS MEDIUM AND THE GAP IS SOURCING. The provider list and the benchmarking both come from the vendor's product page and launch materials rather than from documentation, and no model configuration reference, provider list page or bring-your-own-key statement was reached across two passes. No automatic runtime routing between models is documented, which this axis does not require. One consistency note: local execution via OLLAMA is credited here as model control, and the deployment implication is graded on Dep, not counted twice.

P>F, and the cause of the July error is a documentation-version problem worth recording for the lane. THE OLD DOCS SAY ONE PROVIDER AND THE CURRENT DOCS SAY FIVE. The marketplace knowledge base at confluence.joget.org and dev.joget.org, which the July build cited, reads SINCE OPENAI IS CURRENTLY THE ONLY SUPPORTED MODEL and MORE AI SERVICES WILL BE ADDED IN FUTURE VERSIONS. The current knowledge base at kb.joget.org enumerates the LLM Services as OPENAI, ANTHROPIC, GOOGLE GEMINI, OPENROUTER LLM AND IBM, and the install prerequisite reads AN LLM KEY (ANTHROPIC CLAUDE / OPENAI / CUSTOM MODELS). Both sets of pages are live; one describes an older build. OPENROUTER IS THE ENTRY THAT SETTLES IT rather than the count. OpenRouter is an aggregator fronting hundreds of models across providers, so listing it as an LLM Service means the practical model set is not bounded by Joget's integration work at all. CUSTOM MODELS extends it again to anything the customer runs. BRING-YOUR-OWN-KEY IS THE COMMERCIAL SHAPE and it is the strong form of this axis: the customer supplies the LLM key, holds the provider relationship and pays the inference cost directly, so Joget is indifferent to the choice. A proxy domain setting is documented for environments that route model calls through their own gateway, which is the governance lever an enterprise actually needs. MODEL IS SET PER AGENT in the Define Agent section, so different agents in one application can run on different providers. WHAT THIS DOES NOT INCLUDE is automatic routing between models at runtime, which this axis does not require. The roadmap phrasing in the older pages is now shipped, so nothing here rests on future tense.

Stands at P, and the reasoning needs stating carefully because the strongest-looking fact here points the wrong way. THE UNIVERSAL CONNECTOR IS NOT EXTENSIBILITY. Uploading an OpenAPI spec, or connecting via REST or GraphQL, so the platform auto-discovers permitted actions is EverWorker reaching into the customer's systems. Under the standing one-line test, the vendor's product calling out is integration; extensibility is the customer's systems calling in. That fact is credited on Int and is deliberately not counted here, which is the same correction applied to WhatsMCP on wassist and to MCP on agentx today. WHAT ACTUALLY SUPPORTS PARTIAL is thinner: the platform is reported to expose a simple API. That is a real outbound surface and it is why this is not None. WHY IT DOES NOT REACH FULL, and Mike's 30 August bar is the test: the platform must be callable from outside through a DOCUMENTED API or SDK. No API reference, endpoint list, authentication documentation, SDK package, developer portal or MCP server was reached across two passes, and the simple-API claim rests on trade coverage rather than on a developer surface a reader can inspect. Nothing documents how a customer would invoke a worker from their own code. THE ASYMMETRY IS CONSISTENT WITH THE WHOLE RECORD and worth carrying: this vendor documents inbound reach in unusual detail, spec-driven connectors, a knowledge engine, an agent browser, and documents its outbound developer surface barely at all. That is coherent for a company selling to non-technical business leaders through a services-led motion, but it means a buyer wanting to embed a worker in their own product has little to go on. A developer portal is the first check for this cell at lane close.

Stands at F, re-based off the general plugin claim and onto specifics, with one fact moved off it. THE PLATFORM IS CALLABLE FROM OUTSIDE, which is Mike's 30 August bar. Joget exposes APIs for integration as a core platform property, and the AI Agent Builder is itself distributed as a plugin JAR downloaded from the Joget Marketplace and uploaded through Manage Plugins. That installation path is the extensibility architecture demonstrating itself: the agent capability is not built into the core, it is a plugin like anything a customer would write. THE OPEN-SOURCE CORE IS THE STRONGEST FORM OF THIS AXIS and few records in this lane have it. A customer can read the source, extend it, and build their own plugins against documented extension points rather than being limited to a published API surface. The BeanShell Prompt, Tool and Enhancer components extend that into the agent builder itself, so custom code is a first-class component type inside an agent rather than an escape hatch outside it. WHITE LABEL AND OEM DISTRIBUTION corroborate: the platform is sold to be embedded in other people's products, which is only viable if the extension surface is genuinely complete. ONE FACT MOVED OFF THIS CELL. The MCP Client Tool was implicitly part of the extensibility story; it is Joget calling remote MCP servers, which under the one-line test is integration, and it now sits on Int alone. No first-party MCP server exposing Joget agents to an outside assistant was found, which under the ruling does not withhold the grade. What is not documented on the pages reached is an API reference for invoking an agent specifically from outside, as distinct from the platform's general APIs. That is the natural check at lane close.

F>P under the 31 August Eval ruling, which is the case this cell was made for. WHAT THE BENCHMARKING IS FOR IS STATED IN THE BASIS ITSELF: it tests agents against base models SO TEAMS CAN PICK THE MOST COST EFFECTIVE AND ACCURATE LLM FOR EACH JOB. The comparison axis is the model, and the output is a model-selection decision. The ruling is explicit that model accuracy scores produced by a platform whose product is something else are real measurement of the wrong thing, and Partial is the ruled position. THE CLOSE CALL, RECORDED BECAUSE IT IS GENUINELY CLOSE. The mechanism does run the customer's agent and score its output, which is nearer the bar than an AutoML leaderboard. Had the vendor framed it as testing whether an agent behaves correctly across versions, it would clear Full. It does not: the held constant is the task and the variable is the model, so the question answered is which LLM to buy, not whether this agent is right. WHAT WOULD MOVE IT is a retained test set with expected outputs and a comparison across agent versions. Nothing of that kind appears on any surface reached across two passes. CONSISTENCY NOTE, since this record and mobagel now sit at the same grade for the same reason: mobagel's AutoML compares algorithms, EverWorker's benchmarking compares LLMs, and both produce a readable number about the model rather than about the agent. agentx reached Full today on the opposite pattern, LLM-as-a-judge scoring against test datasets tracked across agent versions. Those three together are a clean illustration of where the ruled line falls. One adjacent fact recorded and not credited: the vendor's blog references converting artifacts into reusable playbooks including EVALUATION SUITES. That is buyer guidance about running an AI programme, not a documented product feature, and it comes from the same blog series that produced the SOC 2 and approvals traps on this record.

Stands at P under the 31 August Eval bar, and the reasoning is the same one applied across this batch. THE PREVIEW TAB IS DOCUMENTED AS PART OF THE BUILD LOOP: TEST YOUR AGENT IN THE PREVIEW TAB, positioned as the step between configuring tasks and saving the agent into a workflow. It lets a builder simulate scenarios, watch the task sequence execute, and refine prompts, tools and enhancers before the agent reaches a live process. For a platform where a published agent becomes a step in a running business process, testing before that point is the right gate and it is in the documented path rather than optional. WHY IT DOES NOT REACH FULL. The ruled bar asks for a result the customer can read AND COMPARE about the agent's behaviour on their own work. Preview shows what the agent does on one input; nothing scores it, retains scenarios as a test set, records expected outputs, or lets a builder establish whether a change to a prompt or tool improved behaviour rather than merely altering it. This is a quality gate with no readable result, which is the ruled definition of Partial. THE DEBUGGING LIMB IS BETTER SERVED THAN THE TESTING ONE, and the gap between them is narrower here than the grade suggests. Because the task model is explicit, prompt then tool then enhancer, a builder watching a preview run can see which step produced a bad output rather than inferring it from a final answer. The Agent Execution Audit Trail credited on Obs extends that visibility into production. WHAT WOULD MOVE THIS is a saved test set with expected outputs and a scored comparison across agent versions. The platform's own data layer would make that straightforward to build, which is why the absence reads as not yet shipped rather than architecturally hard.

N>F. The July basis said agents access the web through integrations and no browser use is documented. The vendor's own launch announcement documents a browser, and names it. THE SENTENCE IS EXPLICIT AND THE STRUCTURE OF IT IS WHAT DECIDES THE GRADE: the platform can INTERACT WITH AND EXTRACT DATA FROM ANY APPLICATION WITH AN API, WHETHER A THIRD-PARTY COMMERCIAL APPLICATION OR IN-HOUSE CUSTOM APP, OR FROM ANY WEBSITE VIA OUR AI AGENT BROWSER. THE WORD OR IS THE AXIS DEFINITION IN MINIATURE. The vendor presents two routes and separates them by whether a programmatic interface exists: API where there is one, AI AGENT BROWSER where there is not. Comp is non-zero precisely when an agent operates software because no programmatic interface is available, and this is a vendor drawing that same line itself and shipping a named component for the second case. INTERACT WITH IS OPERATION, NOT RETRIEVAL, which is the distinction I have used to refuse this axis repeatedly across this batch. Extraction alone would be scraping and would fail; interact with and extract sits on the other side. A named AI AGENT BROWSER is a component, not a capability adjective. THE EVIDENCE TIER IS WORTH NOTING. This comes from the company's own launch announcement, which under section 7 clears the first-party floor as a vendor announcement, and it is the release in which the platform was relaunched, so it describes the current product rather than a legacy one. CONFIDENCE HIGH ON EXISTENCE, and the limits recorded: no documentation of what the browser can do, whether it authenticates, how it handles sign-in, or any scope statement was reached. Desktop or operating-system control is not claimed anywhere; this is browser only. For the lane: this is the third genuine Comp positive found in Agent builder, alongside sema4-ai and kalcend, and like kalcend it sits on a record nobody would have predicted it from.

Stands at N, and confidence rises to high because the full component palette is now enumerated, which turns this from an unfound claim into an observed absence. THE TOOL LIST IS CLOSED AND READABLE: BeanShell, Database Query, Email, Google Web Search, JSON API, Run AI Agent, Mem0 Chat Memory, MCP Client and Pinecone. Every one is a programmatic interface. When a vendor publishes the complete set of tools an agent can be given, absence from that list is genuine evidence rather than a failed search, which is a materially stronger basis than most N cells in this batch. THE NEAR-MISS IS THE GOOGLE WEB SEARCH TOOL and the July basis refused it correctly. Searching the web returns results over an API; it neither navigates nor operates a page. This is the same refusal applied to ai-library's scraping utility and pickaxe's search action today, and the line has been held identically each time. A SECOND NEAR-MISS WORTH NAMING because it is subtler: the BEANSHELL TOOL executes arbitrary custom code. A determined customer could write BeanShell that drives a headless browser, and in that sense the platform can be made to do computer use. That does not earn the grade. The axis measures what the vendor documents as a capability, not what an extensible platform could be extended into, and crediting it would give every open-source platform a Full on this axis by construction. THE ABSENCE IS STRUCTURAL. Joget's premise is that agents act inside governed business processes against the application's own forms, tables and database, and against systems reached through connectors. Driving an interface has no place in that model, and the vendor makes no claim in the territory.