What’s new
- Agents Knowledge Base is now a first-class product area with draft, create, and detail views for building and refining lenses.
- Agents Knowledge Base capabilities now plug into AML and fraud workflows through a reusable subagent. Typology investigation is more complete and returns richer evidence summaries.
- Agents Browser Use recordings now open directly on the target site. Session videos skip the watermarked landing page at the start.
- Agents Reference images from web analysis and browser tasks now render reliably in job results. Access is validated from the job itself instead of browser headers.
- Agents Large PDFs now process in parallel page windows. Extraction is faster and more reliable on long documents.
- Agents Agent planning uses fewer back-and-forth calls. Planning-enabled investigations run with less overhead.
- Platform Agent job lists and metrics load faster at high volume. Roe now uses bounded history windows and feed-style pagination to keep large job histories responsive.
- Platform Job details now show who started the run. Creator information appears in job lists, job detail views, and workflow result pages.
- Platform Large integer IDs in workflow inputs now display exactly as entered. Roe preserves full precision for long alert and case identifiers.
- Integrations Exa deep research now uses the current search flow and supports structured output more consistently. Research and answer calls align better with Exa’s latest API.
- Integrations Merchant risk workflows now use the Resellbot API directly for seller revenue data. This removes a browser step and makes the lookup more direct.
- QA Batch export links now open in the correct organization context. Shared CSV and JSON exports land on the right job view.
- QA Self-hosted Browser Use runs now track cost correctly. Execution traces show more accurate per-step and total usage.
- QA PDF cleanup now removes sticky-note comments without stripping stamps or free-text content. Annotated files preserve the content users expect to keep.
- Dashboard Diagnostics reruns now let you choose the agent version explicitly. New diagnostic runs also default to the current version.
- Dashboard Diagnostic report text now renders in monospace where needed. Inline code, quotes, and markdown are easier to read.
- API Agent context sources are now available through the API and Roe MCP. You can list a source’s read-only operations and call them directly.
- API Public API access to job artifacts now supports result-only retrieval. The endpoint returns just the result payload and strips internal metadata and usage details.
- API Custom MCP connectors now support skill sections. You can upload or paste reusable guidance that gets added to agent instructions when that connector is used.
- API The v1 agent cancel-all endpoint now runs asynchronously. Large cancellation requests return quickly and continue in the background.
- API Roe’s public SDK surface is broader and more complete. More allowlisted operations now have first-class client methods across supported SDKs.
- Security Dataset file access is more tightly scoped to the right organization. Download and image retrieval paths now enforce stronger authorization checks.
- Security Sensitive OTP values are no longer written to logs.
- Security Security patches shipped across Roe services and frontend dependencies. This release includes fixes for newly disclosed vulnerabilities.
- Workflows Workflow inputs are available everywhere in the orchestration canvas. The Workflow Inputs card now stays visible while building mappings and conditions.
- Workflows Workflow orchestrations handle large fanouts and loops more efficiently. Roe now batches preparation work and shards large loop bodies to keep big runs stable.
- Workflows Batch runs can now be exported as CSV from the job sidebar. The export includes original batch inputs, status, and agent results for each row.