v1.5.0

🚀 Notable Features and Improvements

• Dynamo Runtime Intelligence: Automatically infer per-request latency sensitivity from agent profiles and apply runtime hints for cache control, load-aware routing, and priority-aware serving.
• Agent Performance Primitives (APP): Introduce framework-agnostic performance primitives that accelerate graph-based agent frameworks such as LangChain, CrewAI, and Agno with parallel execution, speculative branching, and node-level priority routing.
• LangSmith Native Integration: Observe end-to-end agent execution with native LangSmith tracing, run evaluation experiments, compare outcomes, and manage prompt versions across development and production workflows.
• FastMCP Workflow Publishing: Publish NeMo Agent Toolkit workflows as MCP servers using the FastMCP runtime to simplify MCP-native deployment and integration.

Migration notice: 1.5.0 includes packaging and compatibility refactors (meta-package restructure, eval/profiling package split, and import-path updates). See the Migration Guide.

🚨 Breaking Changes

• Dynamic Inference Headers with Prediction Trie Integration by @dnandakumar-nv in #1483
• improvement(packaging): Packaging Restructure for libraries by @willkill07 in #1512
• fix: Langchain agents should reuse runnable config by @willkill07 in #1604
• Refactor: Split eval/profiler into optional nvidia-nat-eval package by @AnuradhaKaruppiah in #1599
• improvement: nvext.agent_hints and nvext.cache_control clean up by @bbednarski9 in #1648

✨ New Features

• Add evaluator feedback to GA optimizer by @dnandakumar-nv in #1442
• Add similarity scores and threshold filtering to Redis semantic search by @thepatrickchin in #1434
• Add end-to-end custom metadata propagation for observability by @ericevans-nv in #1480
• Expose MCP session ID and add custom headers support for session correlation by @yczhang-nv in #1500
• Preserve workflow state across WebSocket reconnections by @ericevans-nv in #1541
• feat(ci-scripts): utility scripts for license updates and SBOM by @willkill07 in #1548
• Add example of control flow composition with router agent and sequential executor by @thepatrickchin in #1550
• Add support for publishing a NeMo Agent Toolkit workflow as a MCP server via FastMCP3 by @AnuradhaKaruppiah in #1539
• Add dataset store registration and support by @dnandakumar-nv in #1576
• Allow for prompts to be stored in/loaded from files by @pastorsj in #1567
• feat(observability): Cross-Workflow Observability by @willkill07 in #1598
• FastAPI Frontend Refactor with HTTP HITL and OAuth Support by @willkill07 in #1603
• Add support for LangSmith evaluators by @mpenn in #1592
• Add automatic latency sensitivity inference by @dnandakumar-nv in #1618
• RAG Library Mode integration by @ericevans-nv in #1440
• feat: Add HuggingFace Inference API and Embedder providers by @bledden in #1570
• Integrate LangSmith Observability with Evaluation and Optimization by @pastorsj in #1593
• Add experimental nvidia-nat-app Agent Performance Primitives subpackage by @mpenn in #1636
• Add Dynamo Example for Latency Sensitivity Assignment by @dnandakumar-nv in #1634
• Revert unintended changes from PR #1704, preserve nat-ui submodule fix by @ericevans-nv in #1710

🔧 Improvements

• Restore version 1.5 on develop after forward merge by @mnajafian-nv in #1324
• Forward-merge release/1.4 into develop (conflict resolution) by @mnajafian-nv in #1394
• Add OAuth2-Protected MCP Calculator Example by @AnuradhaKaruppiah in #1403
• Forward-merge release/1.4 into develop by @mnajafian-nv in #1453
• Merge release/1.4 into develop by @willkill07 in #1459
• Add use_native_tool_calling option to ReAct agent by @yczhang-nv in #1476
• Add raise_on_parsing_failure option to ReAct agent by @yczhang-nv in #1477
• Enable per-user workflow support in nat eval by @ericevans-nv in #1503
• feat: make tavily internet search tool configurable by @cdgamarose-nv in #1518
• Update nat-ui submodule to latest main by @ericevans-nv in #1551
• chore(pre-commit): update versions in pre-commit; add root-level uv.lock check by @willkill07 in #1553
• Add a new per-user MCP client tool list endpoint by @AnuradhaKaruppiah in #1561
• Add HTTP retry logic and error resilience for workflow evaluation by @ericevans-nv in #1563
• Fix workflow name regression introduced by custom OTEL span naming by @ericevans-nv in #1572
• Add Support for NVExt Annotations and Latency Sensitivity for Dynamo by @dnandakumar-nv in #1575
• chore(ci): add stale action for old Issues/PRs by @willkill07 in #1581
• Update dynamo headers to provide raw integer values by default by @dnandakumar-nv in #1583
• Resolve user ID from JWT or nat-session cookie by @AnuradhaKaruppiah in #1584
• Add support for Weave feedback comments by @thepatrickchin in #1586
• chore(llm-providers): Add env OPENAI_BASE_URL for openai ; unify llm provider configs by @willkill07 in #1577
• HITL prompt timeouts and API error responses by @ericevans-nv in #1591
• Update nat-ui submodule to latest main by @ericevans-nv in #1594
• enh(sbom-licenses): refactor common code; enable multi-version diffs by @willkill07 in #1597
• Populate full connection attributes and payload for HTTP and WebSocket sessions by @ericevans-nv in #1602
• Refactor latency sensitivity to use integers instead of enums. by @dnandakumar-nv in #1601
• chore: update nat-ui submodule by @ericevans-nv in #1606
• Add cache pinning strategy for KV cache with TTL control by @dnandakumar-nv in #1609
• Add type converters for langgraph wrapper nat serve endpoints by @ericevans-nv in #1610
• Add max_sensitivity for latency-based prioritization by @dnandakumar-nv in #1612
• feat(agent): add token-by-token streaming to tool_calling_agent by @MylesShannon in #1595
• chore(deps): upgrade uv.lock deps prior to release; relax dependencies by @willkill07 in #1621
• Allow running pytest from project root by @dagardner-nv in #1622
• Fix warning messages emitted from test_per_user_fastapi_integration.py by @dagardner-nv in #1624
• feature: dynamo integration with nat profiler and prometheus/grafana dashboard by @bbednarski9 in #1486
• Remove ci/release/update_toml_dep.py script by @dagardner-nv in #1646
• CI fix: exclude CHANGELOG.md from pre-commit checks by @bbednarski9 in #1653
• Remove NASSE naming by @ericevans-nv in #1632
• chore: remove all unnecessary docker deployment guides from examples by @willkill07 in #1655
• Remove Profiler Agent Example from the Primary Toolkit Repo by @AnuradhaKaruppiah in #1656
• improvement: nvext.cache_control warning and HiCache for SgLang images by @bbednarski9 in #1658
• Fix multi_frameworks example UnboundLocalError and upgrade default LLM by @ericevans-nv in #1661
• Fixes and improvements for tests by @dagardner-nv in #1659
• improvement(logging): add file logging mode option by @willkill07 in #1651
• Refactor defense and red teaming middleware with pre/post invoke hooks by @ericevans-nv in #1671
• Add a server-side override for the A2A Agent Card URL by @AnuradhaKaruppiah in #1673
• Fix Mem0 metadata validation error and improve auto_memory_wrapper example by @ericevans-nv in #1683
• Add an optional proxy server to map model names for integration testing by @dagardner-nv in #1679
• chore(deps): bump package versions by @willkill07 in #1682
• Fix DynamicFunctionMiddleware builder patching regression by @ericevans-nv in #1691
• Fix incorrect CLI flag in auto_memory_wrapper README by @ericevans-nv in #1692
• Add E2E test for Tool Calling Responses API Agent by @dagardner-nv in #1726
• Fix Milvus connection failures in RAG integration tests by @ericevans-nv in #1724
• feat: add NIM model endpoint health check by @mnajafian-nv in #1716
• feat: add embedder inference check and Slack reporting for model health by @mnajafian-nv in #1736
• Update container tag for the nginx-rewrite-models service by @dagardner-nv in #1740
• Improve the text file ingest E2E test by @dagardner-nv in #1759
• Update uv.lock files by @dagardner-nv in #1762
• Improves FastMCP dev experience and docs by @AnuradhaKaruppiah in #1773
• Observability user experiences fixes by @dnandakumar-nv in #1760
• Updating CHANGELOG and Release Notes by @AnuradhaKaruppiah in #1792

🐛 Bug Fixes

• fix: bump NAT version to 1.5 for packages that were added under release/1.4 by @willkill07 in #1399
• Fix MCP tool validation for nullable optional fields by @AnuradhaKaruppiah in #1507
• fix(serve): ensure a single event loop for python 3.11 by @willkill07 in #1528
• fix: flaky batching processor test by @willkill07 in #1529
• fix(ci): coverage reports should only be for nat code and examples by @willkill07 in #1536
• fix(ci): Fix build_wheels and slack notifications for nightlies by @willkill07 in #1537
• fix(tests): add required deps for some e2e tests; get notebook tests working by @willkill07 in #1538
• Forward-merge release/1.4 into develop by @willkill07 in #1552
• Use relative paths for symlink creation in workflow create command by @thepatrickchin in #1557
• fix(milvus): Fix vector_field config mapping and document_id type by @rmalani-nv in #1555
• Refactor span attribute serialization to use JSON strings by @dnandakumar-nv in #1574
• fix(ci): ensure packaging works in GitLab CI by @willkill07 in #1582
• Fix FastMCP example E2E tests by @AnuradhaKaruppiah in #1580
• fix(gitlab-ci): ensure gitlab artifact upload is configured correctly by @willkill07 in #1588
• fix(ci): ensure stale action has required permissions by @willkill07 in #1589
• fix(mcp): Cache enum classes to prevent validation errors by @bledden in #1564
• fix(tests): prepare for OpenAI endpoint for nightlies; fix failing tests by @willkill07 in #1596
• Refactor call index tracking for prefix predictions by @dnandakumar-nv in #1608
• Fix failures after nvidia-nat-eval isolation by @willkill07 in #1615
• fix(mcp-client): ensure tools are only invoked when available by @willkill07 in #1616
• fix: update openpipe-art accuracy reward logic by @aslanshi in #1623
• fix: Preserve custom dataset fields in workflow output by @bledden in #1628
• fix: Skip output directory cleanup when --skip_workflow is set by @bledden in #1627
• fix: Pass request_timeout through to OpenAI/Azure LLM clients by @bledden in #1626
• fix: Filter empty LLM responses from ReAct retry scratchpad by @bledden in #1629
• Fix auth callback trace and update test scripts by @AnuradhaKaruppiah in #1633
• fix(http-hitl-oauth): fix streaming and default configuration values by @willkill07 in #1641
• Add validation alias for nvext_max_sensitivity in DynamoLLM by @dnandakumar-nv in #1657
• Treat explicit null defaults as nullable in MCP schema translation by @AnuradhaKaruppiah in #1665
• Fix Unicode escape sequences showing in console workflow output by @yczhang-nv in #1664
• Improve ReAct tool input parsing for Python-style Action Input literals by @AnuradhaKaruppiah in #1666
• Add missing dependency for nvidia-nat-opentelemetry to nvidia-nat-langchain by @dagardner-nv in #1670
• Fix ReAct agent parsing failures with reasoning models (<think> tags) by @yczhang-nv in #1667
• Simplify the example questions to bypass priv levels by @AnuradhaKaruppiah in #1672
• Fix thought matching issue in ReAct agent with the Llama-3.1-Nemotron-Nano-4B-v1.1 model by @dagardner-nv in #1675
• chore(deps): provide upper-bound for starlette; bump grpcio versions by @willkill07 in #1669
• Fix setting the openai base url for llama index by @dagardner-nv in #1686
• examples(mcp): make example more robust to LLM hallucination by @willkill07 in #1695
• Fix mixture of agent example from reaching GRAPH_RECURSION_LIMIT by @dagardner-nv in #1697
• fix: handle GraphRecursionError gracefully in tool_calling_agent by @mnajafian-nv in #1705
• Update nat-ui submodule to include conversation state fix by @ericevans-nv in #1704
• Fix alert_triage_agent empty reports in offline mode (#1699) by @mnajafian-nv in #1703
• fix(notebooks): add missing nat workflow reinstall before nat run by @mnajafian-nv in #1713
• fix(simple-web-query): harden tool description; disable thinking by @willkill07 in #1722
• Fix pydantic model validation for nvext hints by @dnandakumar-nv in #1723
• Pin nvidia-nat-ragaai to setuptools v81 by @dagardner-nv in #1730
• Fix Strands integration tests by @dagardner-nv in #1731
• fix: replace llama-3.1-405b model in email phishing analyzer by @mnajafian-nv in #1712
• Fix alert triage agent: switch to nemotron-3-nano model and improve prompts by @hsin-c in #1750
• fix(ci): sanitize sbom license response by @willkill07 in #1763
• fix: replace deprecated mistral-nemo-12b and fix reasoning agent tool discovery by @mnajafian-nv in #1781
• fix: toplevel pyproject.toml should specify tool.uv.managed=true by @willkill07 in #1783

📝 Documentation Updates

• Fix typo in documentation for uv sync command by @thepatrickchin in #1542
• docs: 1.5 migration guide for packaging by @willkill07 in #1625
• chore(docs): add GitHub Issues/PRs to Linkcheck ignorelist by @willkill07 in #1642
• Ensure that we allways spell vLLM with the same casing that the proje… by @dagardner-nv in #1644
• Document the need to set the NVIDIA_API_KEY in the Redis example by @dagardner-nv in #1678
• Add documentation for langsmith evaluators by @pastorsj in #1643
• Organize alternate source/package install commands into tabs by @dagardner-nv in #1737
• Cleanup vale vocabulary by @dagardner-nv in #1745
• Fix out of date documentation references to AIQContext by @dagardner-nv in #1795

New Contributors

• @bledden made their first contribution in #1564
• @pastorsj made their first contribution in #1567

Full Changelog: v1.4.1...v1.5.0