Hey Community,
I went through the September 2026 release notes and the FUSION 2026 announcements and pulled out the items that need action, not just awareness. Sharing in case it saves someone a surprise in production. Corrections and first-hand experience are very welcome!
Note: Everything below comes from public UiPath docs and announcements. Most of it applies to Automation Cloud, so check availability for your deployment type before planning around it.
1. LLM model removals (action needed)
| Model | Vendor | Removal date | Replacement |
|---|---|---|---|
gpt-4o-2024-05-13 |
Azure OpenAI | Sept 30, 2026 | gpt-5.1-2025-11-13 |
anthropic.claude-sonnet-4-20250514-v1:0 |
Amazon Bedrock | Oct 13, 2026 | anthropic.claude-sonnet-4-6 |
What happens on the removal date:
- UiPath-managed LLMs switch to the replacement automatically. There’s no outage, but outputs can change, so re-run your evaluations.
- BYO configurations that reference the removed model are invalidated. If they aren’t updated, UiPath falls back to UiPath-managed LLMs as a last resort.
Suggested checklist:
- Use Model impact preview to find which products and processes depend on the model
- Review your BYO configurations in AI Trust Layer
- Test the replacement in a non-production environment and compare outputs
- Update BYO configs before the removal date
Heads-up: Claude Sonnet 4.6, the Bedrock replacement above, is already deprecated for Autopilot, where Claude Sonnet 5 replaces it. Check the per-product lifecycle, not just the platform-wide table.
Sources:
2. Agents
- Agents in Flow (GA): configure agents directly in a Flow, connect them through Flow variables, and chain multiple agents without a separate publishing step.
- Data Fabric entities as agent context: structured, current data for agents that respects existing access rules.
- Coding Agents (GA): build agents, Maestro Flows, API workflows, apps, and human approval tasks with your coding agent of choice. Treat generated workflows like any other code: review, test, and gate them before deployment.
Sources:
3. Maestro and Orchestrator
- Maestro is now two products: Maestro Orchestrate for long-running, cross-team processes, and Maestro Automate (GA) for high-volume, shorter-running tasks such as agent and API workflows.
- New BPMN patterns: reusable structures for human review, approval chains, and failure handling.
- Environment variables for standard RPA processes, with secret asset references. Requires Robot 25.10 or later.
- Credential stores can now be restricted to selected folders.
- Queue webhooks: events may arrive batched during bulk uploads. If your consumer assumes one item per delivery, update it to handle multiple.
Sources:
4. Worth watching
- Cartographer (GA) and the Map of Work: drafts a map of how work gets done from documents, systems, and people, with each fact tied to a source and a verifier. A Decision Ledger that records who decided what, on what evidence, is in preview.
- UiPath + BDO: announced agentic accelerators for ITGC and ITAC controls testing, covering access reviews, change-management evidence, and exception reporting. This is a partnership announcement, not a GA product.
Sources:
Questions for the community
- Has anyone moved a BYO config from gpt-4o to gpt-5.1 yet? Did you see prompt or output differences worth sharing?
- Did the batched queue webhook events affect your consumers? How did you adapt?
- Is anyone running Agents in Flow in production? How are you handling review and testing of agent steps?
Thanks, and please add anything I missed!