FLOWX.AI vs Obin AI (2026)

Connectors do the integration work. A connector is a microservice the customer builds and deploys in any technology, acting as an anti corruption layer over an external system, responsible for data transformation between formats and for enrichment such as flags and tracing identifiers.

Building one means configuring a Kafka consumer and producer, naming the in and out topics against the pattern the engine listens on, setting consumer threads against partition counts, defining the incoming and outgoing DTOs and implementing the business logic with the process instance UUID as the message key, with optional health checks, and FlowX publishes a quickstart repository and a currency exchange example connector.

Connectors are invoked from the Process Designer through Send Message Task and Receive Message Task nodes, and the engine captures the reply into process variables.

The platform also connects to named systems, including the Jack Henry, FIS, Finastra and Temenos core banking platforms, COBOL mainframes, transportation management systems, custom in house applications and RPA tools, plus databases, REST APIs, files over FTP, SFTP, Amazon S3 and Azure Blob, email, webhooks and AI tools over the Model Context Protocol, leaving the systems of record in place rather than replacing them.

Sourcedocs.flowx.ai 5.9 building a connector and integrations overview pages, flowx.ai platform integrations pageread 2026-10-25

Agents act beyond reading. The loan notice agent posts updates after validating a notice against the credit agreement, deal economics are serviced live in the fund's own models, and the platform writes an audit trail into the firm's existing audit log infrastructure. No system of record, ERP, CRM, data provider or document source is named, there is no connector catalog, and there is no stated custom tool or authenticated action framework. Running inside the firm's environment on the firm's data says where the work happens, not how it reaches outside systems.

SourceObin AI, obin.ai/solutions and obin.ai/platformread 2026-09-07

FlowX processes are built from BPMN blocks, including exclusive (XOR) gateways that route each instance down one path on evaluated conditions and parallel gateways for concurrent branches. A Custom Agent node and BPMN integration place AI agents inside processes, and Agent Builder offers typed AI nodes such as intent classification and context retrieval.

Sourcedocs.flowx.ai 5.9 exclusive gateway, parallel gateway, Custom Agent node and Agent Builder pagesread 2026-09-30

Agents run defined workflows end to end and do not just assist at the task level. First pass diligence runs from source documents through to an investment committee memo, and the loan notice agent reads each notice, validates it against the credit agreement and deal record, posts updates and escalates exceptions.

The platform's own lifecycle is configure, prove, then operate, with every intermediate step of an agent's reasoning validated and a failed step re-reasoned and fixed. Multiple named workflows run in parallel across underwriting, monitoring and reporting under one control plane that manages uptime, throughput and resources.

SourceObin AI, obin.ai/solutions and obin.ai/platformread 2026-09-07

Scheduled processes automate instance creation on dates, times or intervals through Start Timer Event nodes, with timer expressions in cron syntax or ISO 8601 durations, and timer start, intermediate and boundary events are each covered separately. Incoming Webhooks start a process from an external HTTP POST or resume a running instance, secured with API keys and running on events with no polling.

Email Trigger connects to IMAP servers to monitor mailboxes, and Message Start Event starts an instance on a message, incoming email or incoming webhook. Four REST endpoints start processes, including by workspace and process name. Manage Triggers provides a central place to view, activate and deactivate event based triggers. A Chat UI component supports agent conversations with end users, with client SDKs for web, iOS and Android.

Sourcedocs.flowx.ai v5.9 documentation index reached via the published /llms.txtread 2026-09-01

Work arrives in two ways. In the inbound queue, a named customer's agent processes more than a million loan notices a year, reading each notice as it arrives, validating it and escalating the exceptions, so document arrival drives the work, not user invocation.

On state and threshold events, agents continuously track covenants, performance and market signals across the portfolio and surface converging risk signals for review, with the thresholds and escalation paths set per agent at configuration, and an insurance deployment keeps production metrics and goal attainment visible in real time. Evaluations also run continuously against production traffic.

SourceObin AI, obin.ai/solutions and obin.ai/platformread 2026-09-07

The Knowledge Base integration ingests documents into Knowledge Bases that agents query during workflow execution, using vector embeddings with hybrid, semantic or keyword search, multiple stores, relevance scoring, append, replace and delete operations with an audit trail, and entries versioned with the app; companion pages cover managing stores, using Knowledge Bases in workflows and testing queries.

Sourcedocs.flowx.ai 5.9 Knowledge Base integration overviewread 2026-09-30

A maintained structure sits over the customer's own knowledge. The firm's contracts, credit policies, workflow documentation and the historical decisions its best people learned from are ingested, and its policies, risk thresholds and decision frameworks are encoded into a foundational layer that every Robin inherits.

Agents draw cross asset signals and comparables from prior deals pulled from disparate sources, with every recommendation traceable to its source, and the loan notice agent validates each notice against the credit agreement and the deal record. The layer persists across agents and deals, scales past any one run and stays queryable.

SourceObin AI, obin.ai/platform and obin.ai/solutionsread 2026-09-07

Session memory lets AI nodes in chat driven workflows see earlier messages from the current conversation. A builder turns it on per Custom Agent or Intent Classification node with the Use conversation memory toggle, which is off by default, and a Memory tab in the workflow console shows the context that was sent to the model. That built in memory is tied to the chat session ID and is not shared across sessions or users.

For memory that carries from one conversation to the next, FlowX publishes a cross session pattern. Each user's preferences, decisions and key facts are kept as a persistent profile in a dedicated FlowX Database data source keyed by user ID, and a summary of that profile is injected at the start of every new session so the agent can draw on earlier interactions. The profile store is a data source the customer creates and manages in FlowX Database, separate from the business data its processes run on.

Sourcedocs.flowx.ai 5.9 memory capabilities and session state management patternread 2026-10-25

Obin describes institutional knowledge that compounds and persists when people rotate off teams, but that resolves into two other things. The durable part is the foundational layer of encoded policies, risk thresholds and decision frameworks that every Robin inherits, which grounds the agents and does not serve as memory.

The improving part is an agent that improves over time by continuously learning from user corrections and actual outcomes, which is learning absorbed into the agent, driven by linking agent traces to outcomes in testing. There is no stated memory layer with a scope and lifetime, and nothing says where memory lives or how a customer firm deletes it for one of its own clients.

SourceObin AI, obin.ai/platformread 2026-09-07

The hybrid AI and business rules pattern includes a runtime human gate as a named variation, with approval gates placed after AI steps so a reviewer sees the AI output and can correct it before it feeds the next step. The gates are built with User Task nodes in the parent BPMN process and are recommended for initial deployment and for high stakes decisions.

A confidence weighted variation feeds the AI confidence score into the business rules, so low confidence extractions tighten thresholds or flag the result for manual review. The alternating AI and script structure leaves a traceable record of which extraction produced which input value, which formula computed which result and which classification drove which filter.

Above the runtime, the FlowX.AI 5 architecture gates agent self modification of the platform, with agents proposing workflow changes and builder agents implementing them only after a designated operator approves. Observatory adds a policy engine evaluated against runs with Guards checking content before it reaches users, and tracks human in the loop and override mechanisms as EU AI Act controls with evidence attached.

Sourcedocs.flowx.ai 5.9 hybrid AI and business rules pattern and Observatory pages, vendor submitted correction verifiedread 2026-09-09

Gating is configurable per agent. Subject matter experts set the thresholds, escalation paths and guardrails for each Robin during configuration, role based permissions control who can configure, ship or suspend an agent, and safety switches are built into all agents, which the co-founder describes as a kill switch on any agentic workflow that strays from the firm's guardrails. In production the loan notice agent escalates exceptions and does not decide them, and the asset manager deployment surfaces converging risk signals for human review with the team retaining full control.

SourceObin AI, obin.ai/platform and obin.ai/solutionsread 2026-09-07

FlowX.AI operates an information security management system certified against ISO/IEC 27001, and an independent SOC 2 Type I examination confirmed its controls were suitably designed as of 17 June 2026 against the AICPA Trust Services Criteria for Security, Availability and Confidentiality. The SOC 2 Type II examination over an operating period is in progress, personal data is processed under GDPR, and the SOC 2 report and ISO 27001 certificate details are available under NDA through the contact form.

Platform security has five layers. Identity is delegated to an external provider, with step by step Keycloak and Microsoft Entra ID configuration. Role based access control works at organization, workspace and project level with a published roles and permissions matrix, and separate runtime authorization for each project governs who reaches a published solution through project roles, end user groups and solution sharing.

A dedicated Audit service centralizes platform events, and a Personal Information Guard detects and redacts personal data before it reaches a model and restores it afterwards. Swimlanes, business filters that restrict process instances by a business value, and permission based expressions that control UI visibility by role round out the controls, while Observatory's compliance mappings cover the customer's own AI apps.

Sourceflowx.ai trust page, docs.flowx.ai 5.9 Securityread 2026-09-11

Role based permissions govern who can configure, ship or suspend a Robin, safety switches are built into every agent, thresholds, escalation paths and guardrails are set per agent with every configuration decision traceable, the audit trail is written into the firm's own existing audit log infrastructure so compliance can review any agent at any time, and the platform runs inside the firm's governed environment on the firm's data. No SOC 2, SOC 1, ISO 27001 or other certification is published, and there is no trust or security page and no trust subdomain.

SourceObin AI, obin.ai/platformread 2026-09-07

Observatory is FlowX's layer for watching AI agents at work. An SDK decorates agents, chains and tools and streams events to the Observatory API, which stores telemetry in PostgreSQL and optionally routes content through Observatory Guards for safety checks. A run is one execution trace, and an event is a log entry inside it. Four pillars read from that same store. Observability gives real time tracing of every agent, chain and tool call with cost and latency analytics, drift detection and threshold alerts.

Governance gives a policy engine evaluated against runs, a six dimensional risk score per app, evidence collection and assessments, with apps registered in an AI registry. Compliance maps EU AI Act, NIST AI RMF and ISO 42001 controls to the runtime, with eighteen EU AI Act requirements scoped by risk tier, each resolving to met, partial, gap or out of scope, and a one click audit pack export returning per requirement evidence with timestamps, telemetry summaries and a gap analysis.

ROI covers risk adjusted return, payback, NPV and Monte Carlo sensitivity per agent and project. Retention is configurable, with a seven year setting offered for the EU AI Act record keeping requirement, and Observatory can be self hosted as well as used as a managed service. These compliance mappings cover the customer's own AI estate.

Sourcedocs.flowx.ai 5.9 Observatory overview and EU AI Act compliance pages, vendor submitted correction verifiedread 2026-09-09

Any action a Robin takes can be traced back to the inputs, policies and calculations behind it, so that a regulator asking questions gets a defensible answer, and the audit trail is written into the firm's existing audit log infrastructure for compliance to review at any time.

Every configuration decision is traceable, agent traces are linked to user actions and eventual outcomes, and the operate view carries proof of the work performed and the outcome it produced alongside uptime and throughput. Every recommendation is traceable to its source. So the record reconstructs why an agent acted, not only that it did.

SourceObin AI, obin.ai/platform and obin.ai/solutionsread 2026-09-07

FlowX can run in the customer's own infrastructure as a containerized platform on Kubernetes, with setup guides for the FlowX Engine, Application Manager, Runtime Manager, Kafka authentication, Redis, ingress routing and authorization, and SaaS terms are published beside license terms. No managed cloud region selection is described.

Sourcedocs.flowx.ai 5.9 setup guides and flowx.ai legal footerread 2026-09-30

The customer environment is the deployment model. Agents are configured, proved and operated inside the firm's cloud, on the firm's data, within its governed environment, and Robin agents run critical workflows today inside each firm's own secure environment, with a named customer's portfolio risk observability running in the firm's environment. Under Obin's open, no lock in architecture, institutions also retain ownership of models, data and IP.

SourceObin AI, obin.ai/platform and obin.ai/solutionsread 2026-09-07

FlowX lists 246 pre designed agents by industry and business area, 80 for banking, 69 for insurance and 97 for logistics, each with a one line function and a deployment effort estimate of low, medium or high driven mainly by how many systems it integrates with.

FlowX describes catalog agents as blueprints rather than shipped components, each a pre designed combination of platform capabilities built with Agent Builder nodes and integration workflows, and every one is a variation of a small set of recurring agent typologies, each with its own page naming the nodes that implement it.

Agents are grouped by business area into multi agent stacks, nine covering retail mortgage underwriting and seven covering commercial onboarding, coordinated with the multi agent orchestration patterns, with a fraud investigation stack of seven (alert triage, false positive screening, evidence compilation, timeline generation, case narrative, SAR compilation, call transcription) and a garnishment processing stack of six among the named compositions.

The stacks page lists 50 stacks in all, 11 for banking, 9 for insurance and 30 for logistics. The same catalog can be browsed on the FlowX site with filters and detail views under a heading of 246 agents, while the homepage and about page round the count to 220 plus.

Sourcedocs.flowx.ai 5.9 agent catalog, flowx.ai agent catalog and agent stacks pagesread 2026-10-25

Obin names a catalog. A customer builds an agentic workflow from scratch or starts from a pre-configured template in the catalog, adapted to the firm's policies, thresholds and analytical frameworks before it ships, so teams start their regular workflows from templates and do not build each agent from a blank page. Named workflows ship across underwriting, portfolio monitoring, LP reporting and institutional knowledge for asset management, with further sets for wealth management and insurance, and each template does its own job when selected.

SourceObin AI, obin.ai/solutions and obin.ai/platformread 2026-09-07

Customers configure model providers themselves under Organization Settings, AI Settings, Model Providers. They add a provider from a catalog, supply their own API key or use the FlowX managed key, test the connection against the provider API, and then whitelist which discovered models are available downstream, grouped by capability with context window and indicative pricing shown for each model.

Defaults and Fallbacks assigns a default and an optional fallback model for each AI capability (text generation, image understanding, embeddings, document and OCR, developer agents, audio), set separately for Sandbox, Staging and Production workspaces. Workflow designers override the model on individual AI nodes using a portable provider and model identifier with an optional fallback provider and model, and the model is chosen from the node override first, then the workspace default, and otherwise the run errors.

Supported providers are OpenAI with FlowX managed or bring your own key and EU, US and Global base URL presets, Azure OpenAI as bring your own key against the customer's own deployment endpoint, Mistral as bring your own key across its chat, vision, embedding and OCR lines, and any OpenAI compatible endpoint through a custom base URL. Two named permissions gate reading and editing the configuration. New providers in the catalog arrive on SaaS first, at release 5.10, and reach self hosted deployments in the next LTS release family.

Sourcedocs.flowx.ai 5.9 AI providers and model configuration, vendor submitted correction verifiedread 2026-09-09

The model behind the agents goes unnamed, and there is no stated customer or admin selection of it. Agents are configured, proved and operated on the firm's data without losing the firm's IP, and Obin's ownership language concerns the customer's models, data and intellectual property, not a choice of foundation model. A model agnostic architecture claim is a category claim, not a selection surface.

SourceObin AI, obin.ai/platform and obin.ai/solutionsread 2026-09-07

FlowX publishes SDKs for React, React Native, Angular, iOS and Android for embedding FlowX processes in client applications, API reference pages for platform resources such as storage buckets and objects, and an MCP API reference for managing MCP servers and tools, the last marked as internal paths under review.

Sourcedocs.flowx.ai 5.9 SDKs overview and API referenceread 2026-09-30

There is no stated route for calling Obin from outside, with no API, SDK or MCP server, no developer surface and no documentation site. Obin's open, no lock in architecture is an ownership and licensing posture, not an interface, and the audit trail written into the firm's existing log infrastructure runs outward.

SourceObin AI, obin.ai/platformread 2026-09-07

Evaluations lets customers test their own agents from end to end. A dataset attaches to a specific AI node in a specific workflow and holds test cases of input plus expected output plus optional retrieved context, added by hand or captured from a real node execution in a production run and marked with its source and a back reference.

A catalog of ten built in evaluators covers general quality (correctness, conciseness, hallucination, answer relevance), RAG quality (groundedness, helpfulness, retrieval relevance) and deterministic scorers (exact match, Levenshtein, JSON match against a declared output shape contract), with custom evaluators writable against a supplied scoring prompt and a numeric or binary pass threshold.

An experiment runs the workflow against every test case and scores each test case and evaluator pair through the async evals-judge service, producing aggregate scores per evaluator plus per test case rows showing input, actual output and expected output side by side, the judge's own reasoning, and pass or fail against each threshold.

Every experiment freezes the node configuration at launch, so a baseline run and a run after a change are compared per evaluator and a drop on any evaluator shows a regression between agent versions, and FlowX advises growing the regression set from flagged production executions.

Datasets, test cases and experiments are also exposed as REST endpoints for use from the customer's own tooling, and the judge model is resolved per organization through an EVAL capability an admin can rebind. Custom evaluators, multi dataset experiments and AI improvement suggestions carry later release badges.

Sourcedocs.flowx.ai 5.9 Evaluations, vendor submitted correction verifiedread 2026-09-09

The customer runs an evaluation harness against its own artifacts. Every Robin agent is tested against the firm's own evaluation set, built from the edge cases, ambiguous language and disputed clauses where generic models fail. Every intermediate step of the agent's reasoning is validated with deterministic, repeatable scoring, and a failed step is re-reasoned and fixed.

Evaluations run continuously against production traffic to catch drift and regressions, and each agent is versioned like production software with named releases, deployment history and rollback. Agent traces, user actions and eventual outcomes are linked to improve the agent over time, a controlled post deployment loop, and what is tested is the customer's deployed agent.

SourceObin AI, obin.ai/platformread 2026-09-07

The Browser Automation node is available from FlowX.AI 5.13.0 on SaaS deployments. It drives a real browser to complete tasks on web applications that have no API, including filling multi page forms, navigating portals and retrieving results, as the customer describes the task in natural language and an AI agent plans and executes the browser interactions inside the platform's execution, logging and document storage layers.

Outcomes are checked against the target site's recorded state changes rather than the agent's own success claim, browser sessions persist across workflow steps, and every run is logged through the platform's execution and observability layers. FlowX directs builders to an API integration where one exists and to the read only Web Page Extractor node where only page content is needed.

Underneath, a real Chromium session is driven through the open source browser-use library, with per step logging of action, reasoning and next goal and a screenshot per step, and FlowX has published a 405 run benchmark of nine models over nine browser tasks scored against the fixture site's recorded state. Self hosted deployments receive the node with the next LTS release family. No terminal, desktop or screen scraping automation is claimed.

Sourcedocs.flowx.ai Browser Automation, vendor submitted correction verifiedread 2026-09-23

Agents work over the firm's documents, policies and data inside its own environment and reach outward through ingestion and posted updates, not by operating an interface. There is no stated browser, desktop or remote computer control.

SourceObin AI, obin.ai/platform and obin.ai/solutionsread 2026-09-07