Agent Studio: Release notes

Environment variables can now be defined directly in Digital Worker and Skill settings, making it easier to manage reusable values such as base URLs and other configuration inputs.

Customer benefit: teams can manage environment-specific settings more safely and consistently without hardcoding values into configurations or instructions.

What changed

• UAT and Production workspaces maintain their own environment variable values independently.

• In UAT, users can add, edit, and delete variables.

• In Production, variable names are read-only while values remain editable.

• Environment variables are included in settings visibility and configuration flows during import and export activities.

An @ attribute autocomplete experience is available in the Instructions editor for Digital Workers and Skills.

Customer benefit: users can insert environment variable references more quickly and accurately while creating instructions.

What changed

• Typing @ followed by at least three characters starts a search of configured environment variables.

• Selected values are inserted automatically in the ${env.attributeKey} format.

• Multiple environment variable references can be inserted into the same instructions content.

The same @ attribute search capability is available in Intent-Based Output Format fields used when configuring a State.

Customer benefit: this helps authors create responses and field descriptions faster while keeping environment variable usage consistent.

What changed

• Available in both Attribute Description and Submit Message fields.

• Typing @ followed by at least three characters searches saved environment variables.

• Selected attributes are inserted in the ${env.attributeKey} format and can be used multiple times in the same field.

The IFSAPPS10 connector configuration now supports conditional field rendering for Azure AD authentication.

Customer benefit: admins see the Azure Tenant ID field only when it is relevant, making connector setup clearer and reducing unnecessary configuration fields.

What changed

• The Azure Tenant ID field is shown when the selected identity provider or authentication mode is Azure AD.

• The field remains hidden for other authentication choices where an Azure Tenant ID is not required.

• This improves the IFSAPPS10 connector setup experience by aligning required fields with the selected authentication method.

MetaPrompter is the natural-language interface within Agentic Studio for building and editing agents. Instead of working through configuration screens, users describe what they want an agent to do — for example, “Create an agent that monitors support tickets, classifies urgent ones, and sends a Slack notification to the escalation channel” — and MetaPrompter generates the instructions, tools, connectors and triggers needed, updating a visual graph of the agent as it goes. Changes are always made in a safe UAT environment first and promoted to production through a controlled step.

Key Capabilities

• Coding-agent mode, so MetaPrompter can generate and edit code-level logic in addition to configuration.

• Create a new IFS tool directly from a MetaPrompter conversation, powered by the IFS Intelligence MCP.

• State-based outputs and output attributes, for richer, structured agent responses.

• Create a new Skill directly from chat, using the same natural-language flow used for tools.

• Add system tools to an agent directly from the MetaPrompter conversation.

• Marketplace visibility for tools, skills and components created through MetaPrompter, so they can be reused across workspaces.

Application Builder is a new Agentic Studio capability that extends natural-language creation from a single agent to a complete application. Users describe the application they need, and a meta-agent generates the working application — its logic, its user interface, and its integrations — ready to be hosted on IFS.

Key Capabilities

• Natural-language application generation, powered by a dedicated meta-agent.

• Direct connection to MetaPrompter, so the same conversational experience used to build agents extends to building applications.

• Support for MCP-based integrations within generated applications.

• Hosting of generated applications directly on IFS.

Digital Crew is the platform's next step beyond single-purpose Digital Workers: a coordinated team of Digital Workers that owns an entire process — from first touch to final step — so nothing falls through the cracks between individually automated tasks.

Sample crew workflows are being defined across four domains:

1. Field Service Management: First Time Fix, Proactive Maintenance, SLA Priority, Repeat Visit Prevention, Revenue Recovery

2. Procurement

3. Enterprise Asset Management

4. Finance: Faster Close, Cash Visibility, Touchless Transactions

Ontology is the platform's shared knowledge model — a structured map of business entities, attributes and relationships that MetaPrompter and Application Builder both draw on when creating new agents, tools and applications. Rather than each agent or application defining the business on its own terms, they share a common, evolving understanding of it. This reduces rework, improves consistency across what gets built on the platform, and makes new agents and applications faster to generate accurately.

What is Usage Tracking?

Usage Tracking gives visibility into how Digital Workers are actually being used — every invocation, the outcome it produced, and the underlying model activity behind it — captured consistently across the platform rather than left to each Digital Worker to report on its own.

Rather than building a separate, platform-specific analytics layer, IFS Loops connects this telemetry to Paid.ai, a business engine purpose-built to turn agent activity into usage and billing records. IFS Loops focuses on capturing accurate, complete telemetry from every Digital Worker; Paid.ai turns that telemetry into the usage insight, invoicing and revenue reporting that follow from it.

Summary of Benefits

• Every Digital Worker's usage is captured the same way, at the source, with complete traceability back to the interaction that produced it.

• Usage and billing analytics are handled by Paid.ai, a platform purpose-built for that job, instead of a bespoke build IFS Loops would otherwise have to maintain.

• Activity can be understood in the commercial terms that fit the process — per agent, per action, per workflow, or per outcome.

• Consistent usage data becomes a foundation for governance, audit and commercial reporting alike, not a one-off report.

• Usage and cost are visible directly in IFS Loops Studio, scoped per tenant and per workspace, without needing a separate reporting tool.

Highlights

What shipped:

• Input-screening guardrails for Digital Workers built on NVIDIA NeMo Guardrails, protecting against prompt injection and jailbreak attempts.

• Multi-layer detection: fast regex pattern checks, followed by an LLM-based safety judge (currently evaluating Qwen3Guard-4B).

• Coverage for both direct input (user-typed prompts) and indirect input (documents/invoices the Digital Worker fetches, including OCR'd scanned content).

Why it matters:

• Prevents attempts to override Digital Worker instructions, exfiltrate data, or bypass approval steps (e.g., approving a payment without the required approval workflow).

• Reduces risk from both directly typed attacks and indirect attacks hidden inside supplier/invoice documents — a scenario not well covered by generic industry benchmarks.

Who benefits:

• Builders and operators of IFS Loops Digital Workers who need assurance that ERP-facing agents (invoices, POs, supplier data) are protected from malicious or manipulated input.

New Capabilities

• Guardrail Orchestration:

  • Central decision point that routes all untrusted text (typed or fetched) through the same screening pipeline before it reaches the main LLM.

• Layered Screening Engine:

  • Layer 1 – instant regex rules for obvious attack patterns (e.g., “ignore instructions”).

  • Layer 2/3 – model-based classification for semantic and novel attack detection, now standardized on Qwen3Guard-4B.

• Indirect-Injection Coverage:

  • OCR pipeline added so text hidden inside scanned invoices/images is screened the same as typed text.

• IFS-Tuned Evaluation:

  • Testing extended beyond generic public benchmarks to an IFS procurement/ERP-specific dataset to better reflect real Digital Worker risk (e.g., PO/invoice vocabulary, tool-call abuse).

Known Limitations / In Progress

• Jailbreak-detection accuracy is currently below target on IFS-tuned test cases and is being iterated on; prompt-injection detection is already performing well.

• End-to-end latency of the multi-layer check is being measured and optimized; current runs show multi-second average latency per request.

• Final approach/model selection will be confirmed once IFS-tuned dataset testing concludes.

Better visibility into marketplace adoption and template origins

• Global Marketplace entries now provide clearer signals to help teams decide which templates to use.

• Digital Workers and Skills display total download counts and average user ratings directly on cards and detail views, making it easier to identify popular and well-received templates.

• Imported templates also now show where they came from, such as the Global Marketplace, Local Marketplace, or a ZIP package, and ZIP-based imports start with a clear initial version.

Safer digital worker lifecycle management and improved dashboard collaboration

• Digital worker administration is now more resilient and collaborative.

• The platform prevents deactivation when a digital worker is still actively used in Autopilot, Application Widgets, or Collaboration Widgets, reducing the risk of accidental disruption.

• Dashboard updates can also now be published more deliberately, helping teams share finalized analytics and workspace views with the right users.

Private direct messaging for AI Agents in Microsoft Teams

• Microsoft Teams integration now supports direct 1:1 conversations between users and AI Agents in addition to group channels.

• This enables more targeted Human-in-the-Loop interactions and private follow-ups, while preserving conversation continuity when a user replies back in the bot chat.

Expanded connector support and more flexible channel associations

• Agentic Studio now includes native support for the IFS Model Context Protocol Server, allowing teams to discover and import MCP tools directly from the integrations catalogue with ready-to-use schemas.

• Homepage channel configuration has also been expanded so a single channel can be associated with multiple digital workers, making deployment patterns more flexible.

New Ops role for operational users

• A new Ops role has been introduced to give operational users the right level of access across Agentic Studio without requiring elevated Admin permissions.

• This helps teams separate day-to-day operational responsibilities from full administrative access while keeping cross-environment workflows available to the right users.

Improved canvas viewing and version navigation

• The authoring experience is now smoother when working with complex digital worker workflows and multiple saved versions.

• The Fit View control on the canvas now correctly centers and scales the graph, while the version selector has been refined to make working drafts easier to find and switch to.

More reliable Teams navigation and tool authentication

• Several fixes improve reliability across Teams-based workflows and connected tools.

• The Teams Personal App now handles unavailable default integrations more gracefully, Human-in-the-Loop notification links route more consistently, and invoice generation flows no longer fail because of incorrect authentication handling in valid sessions.

More reliable Teams navigation and tool authentication

• Several fixes improve reliability across Teams-based workflows and connected tools.

• The Teams Personal App now handles unavailable default integrations more gracefully, Human-in-the-Loop notification links route more consistently, and invoice generation flows no longer fail because of incorrect authentication handling in valid sessions.

Super Admin audit visibility for marketplace and import activity

• Additional audit visibility has been introduced for internal governance purposes.

• Super Admin users can now review marketplace interaction history and template movement more effectively, including activity related to downloads, ratings, local marketplace usage, and import source tracking.

Sentiment notifications have been enhanced to provide clearer context about why a score has decreased, including the underlying drivers that contributed to the change.

Customer benefit: teams can understand negative sentiment shifts faster and take action sooner without needing to manually investigate each case in detail.

What changed

• Alerts can now surface more detailed reasons behind a score drop.

• This may include contributing sentiment tags, phrases, or tone indicators that explain the decline.

• Users get more immediate context directly from the notification experience.

The Ask experience has been moved to a more visible position within the Copilot widget so users can enter questions more naturally while staying in context.

Customer benefit: this improves discoverability and encourages faster adoption by making it easier for agents to ask questions without leaving their current workflow.

What changed

• The Ask input is displayed directly below the widget section for easier access.

• The text box remains visible while users browse widget content such as intents, resolutions, insights, and comments.

• Selecting the Ask button places focus into the input so users can begin typing immediately.

Release 0.8 includes additional content-management capabilities across integrations and core services, covering publicly accessible Fluid Topics content, SkillJar connector support, and attachment retrieval for IFS Cloud KB articles.

Customer benefit: teams can bring more external and knowledge-base content into Loops workflows with less manual handling and better connector coverage.

What changed

• Added support for fetching publicly accessible Fluid Topics content that does not require users to sign in or provide additional authorization.

• Implemented SkillJar as a ContentManagement integration, including API configuration, authentication, content fetching, mapping, error handling, and test connection support.

• Added support for pulling Knowledge attachments for IFS Cloud KB articles.

Human-in-the-loop Microsoft Teams messages have been enhanced to support tagging or mentioning users directly in Teams messages.

Customer benefit: teams can draw the right person’s attention to HITL requests faster, improving collaboration and response times in Teams-based workflows.

What changed

• Added support for tagging users in HITL Teams messages.

• Aligns Teams HITL behaviour with the expected Microsoft Teams mention experience.

Template handling has been improved by validating and normalizing Digital Worker and Skill JSON in both export and import flows.

Customer benefit: this reduces template compatibility issues, lowers the chance of import failures, and helps ensure downloaded templates stay aligned with the latest expected schema.

What changed

• Before download, JSON is validated against the latest schema and missing required keys or objects are added automatically where needed.

• Before upload or import, templates are validated again and normalized when possible.

• Existing valid data is preserved while missing defaults are added safely.

• If a template cannot be normalized, users receive a clearer validation error instead of a later-stage import failure.

A fix has been delivered for an issue affecting state-based output rendering when Retrieval was configured as the final tool in the flow.

Customer benefit: output experiences are now more consistent and reliable in retrieval-based scenarios, reducing confusion and unexpected response behavior.

What changed

• Resolved a rendering issue that impacted state-based outputs in workflows ending with Retrieval.

• Improves consistency between configured tool behavior and the final presented output.

An issue affecting responsiveness for the Supplier Order Manager and Supplier Invoicing Manager Digital Workers in Production has been resolved.

Customer benefit: users can once again rely on these Digital Workers to process and respond to queries in live Production use cases.

What changed

• Fixed the issue preventing the affected Digital Workers from responding to user queries in Production.

• Restores expected conversation behavior for these live customer-facing assistants.

Resolved an issue where the Meta API failed when users navigated to tools associated with a Skill on Concept environments.

Customer benefit: users can move through the Skills experience more smoothly without interruptions when reviewing associated tools.

What changed

• Corrected the failure impacting tool loading inside the Skills module.

• Improves reliability when opening Skills that include associated tools.

Fixed an issue where adding the IFS Bot to a Microsoft Teams channel briefly displayed an error message even though the tab was added successfully.

Customer benefit: users now get a cleaner, more trustworthy setup experience when adding the bot in Teams.

What changed

• Removed the transient error message shown during IFS Bot tab creation.

• Tab setup now completes without displaying a false failure message.

This release delivers new and enhanced capabilities for building and managing Digital Worker (DW) templates, deployment, testing, and versioning. The Agent Studio now provides:

• Enhanced Digital Worker build flow

  • Guided build flow for Digital Workers: Create → Instructions → Tools → Skills → Triggers → Settings

  • Instructions as the core control plane: define persona/role, context, guardrails, tool usage, response structure, and multi-step reasoning

  • Tool management is simplified, with connector-based standard tools and custom tool creation as needed

  • Skills as reusable reasoning patterns (prompt, reasoning, and domain skills) to standardize high-quality outputs

• Import/Export of Digital Workers

  • Export as Template workflow with guided checklist (e.g., overview → integrations → branding & metadata → review & export) to make installations predictable and secure

  • Export summary provides an at-a-glance inventory of what’s included (e.g., number of skills, tools, and triggers) and allows editing the system description shown to installers

  • Clear scoping during export: exporting affects the template package only (does not modify the live system)

  • Marketplace import includes an install readiness checklist and surfaces required integrations/connectors that must be configured before the agent can be installed

• New marketplace experience to discover industrial AI agents and deploy enterprise-grade solutions

• Browse and filter by categories (e.g., Quality Control, Predictive Maintenance, Automation, Supply Chain, Safety & Compliance, Process Optimization, Operations, Logistics)

• Search across agents, categories, and features; preview and view agent details before deployment

• Agent cards highlight key signals such as ratings, last-updated date, usage/downloads, and badges (e.g., Trending, Popular, New, Stable)

• Scope (Q1 2026): available for editing existing Agents only

• Interactive prompt designer that generates structured instructions

• Enforces state-based output formats for consistent, predictable behavior

• Supports iterative instruction updates to accommodate customer-specific rule changes without code

• Dual-path memory initialization: combine continuous entity-specific learning with optional user-defined Rule Books for bootstrapping

• Captures learnings from HITL feedback to evolve instructions and long-term memory over time

• Designed to improve resilience to variable inputs (e.g., changing document formats and incomplete data)

• Digital Worker Activity overview with trend charts and KPI tiles (e.g., total events, HITL conversations, total tests)

• Current Activity snapshot (active Digital Workers, active conversations)

• Flow monitoring with completion visibility (Digital Worker flows and close rate)

• Leaderboards for operational triage (Top Digital Workers; HITL conversations by agent)

• Governance/quality surfaces (Eval failures) and cost visibility (token usage by agent)

• Dataset-based eval runs with scorecards across key dimensions (Accuracy, Relevance, Helpfulness), plus operational metrics (tests run, success rate, failures, average latency)

• Per-test-case drill-down: view the agent plan, conversation transcript, and detailed rubric scoring with reasoning to support faster debugging and iteration

• Run history tracking to compare results over time and spot regressions

• Organize tests into datasets with individual test cases and pass/fail status tracking

• Run history captures when an eval was executed and by whom, making it easier to compare runs and detect regressions

• Performance signals included alongside quality (e.g., per-test latency and average runtime) to catch speed/timeout failures early

• Deterministic intent based agent output

• Predictable responses for machine interaction (schema/state-based output normalization)

• MS Teams Adaptive cards support

• None