Python agent release notesRSS

Notes

This release of the Python agent fixes issues with signed headers in botocore (SigV4).

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

Bugs

• Fix issues with signed headers in botocore (SigV4)

  • The previous change to botocore distributed tracing headers has been reverted to avoid duplicate headers being sent.
  • Previously any distributed tracing headers that were generated by other tracing libraries could cause issues with instrumentation in botocore due to header signing. This has been fixed by disabling signing for these headers within botocore.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds support for OpenSearch, Microsoft Azure SQL, new Kinesis methods, and a new browser_monitoring.version setting. Additionally, improvements have been made to Anthropic, MCP, and Google ADK instrumentation. Issues with signed headers in botocore (SigV4) and Gemini tools parsing have also been fixed for this release.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

New Features

• New instrumentation for opensearch-py

  • Instrumentation was added for the official OpenSearch database connector.

• New instrumentation for mssql-python

  • Instrumentation was added for Microsoft Azure SQL.

• New instrumentation for new Kinesis methods

  • The following methods are now supported in Kinesis:
    • create_channel
    • delete_channel
    • describe_channel
    • list_channels
    • update_channel

• New setting for browser_monitoring.version

  • The new setting browser_monitoring.version enables users to specify specific versions of our Browser Monitoring Agent to be injected.
  • See the list of Supported Browser Agent Versions to choose from.

Improved Features

• Improved support for anthropic.

  • Improved parsing of request parameters for the Anthropic SDK on versions 1.0.0 and above.
  • New parsing for responses via Messages.with_raw_response.create enabling improved compatibility with LangChain.

• Support for mcp v2.0.0 and above.

  • Instrumentation has been updated following the major refactor done in mcp v2.0.0 and should once again work as expected.

• Support for Google ADK v2.9.0 and above.

  • Instrumentation has been updated for Google ADK on versions 2.9.0 and above.

Bugs

• Fix issues with signed headers in botocore (SigV4)

  • Previously any distributed tracing headers that were generated by other tracing libraries could cause issues with instrumentation in botocore due to header signing. This has been fixed, and distributed tracing headers are now mutated before signing.

• Fix Gemini tools parsing

  • Previously, some instances of tools created with Gemini would fail to parse and would silently fail, resulting in no response for that specific request. This has been fixed.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds support for Azure CosmosDB, adds instrumentation for HTTPX2 and new LangChain/LangGraph streaming events APIs, adds support for newer OpenTelemetry database semantic conventions, and resolves several issues including incomplete package reporting, a package lookup performance penalty, and a MySQL database name reporting bug.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

New features

• Added support for Azure CosmosDB

  • The Python agent now instruments Azure Cosmos DB client operations, reporting datastore metrics and traces for database, container, and user operations performed through the azure-cosmos SDK.

• Added support for HTTPX2

  • The Python agent now supports HTTPX2, a fork of HTTPX actively maintained by the Pydantic team. External call instrumentation for HTTPX2 mirrors the agent's existing HTTPX support for both synchronous and asynchronous clients.

• Added LangChain/LangGraph streaming events instrumentation

  • Added astream_events instrumentation support for LangChain agents.
  • Added stream_events/astream_events instrumentation for LangChain/LangGraph v3, covering iteration over GraphRunStream objects.

• Added support for newer OpenTelemetry database semantic conventions

  • When running the Python agent alongside OpenTelemetry (Hybrid Agent mode), the agent now reads both the legacy and newer OpenTelemetry database span attribute names (db.name/db.namespace, db.statement/db.query.text, db.mongodb.collection/db.collection.name), so database tracing continues to work correctly regardless of which semantic convention version the OpenTelemetry instrumentation library emits.

Bugs

• Fixed incomplete package reporting to New Relic

  • Previously, the Python agent could omit some installed packages from the Environment data reported to New Relic if module discovery took longer than the fixed time window allotted during the slow harvest cycle, and the package list wasn't refreshed after an agent restart. Package capture now runs during the fast harvest cycle and is fully re-captured on agent shutdown/restart, so the reported package list is complete.

• Reduced overhead of package version lookups

  • Previously, looking up installed package versions rescanned the filesystem for every package on every lookup, incurring an O(n^2) time penalty that scaled with the number of installed packages during agent initialization. The result of importlib.metadata.packages_distributions() is now cached after the first lookup to avoid this overhead.

• Fixed get_linking_metadata to properly pass through the specified application

  • Previously, calling newrelic.agent.get_linking_metadata(application=...) ignored the application argument and always returned linking metadata for the default application. This affected processes reporting to more than one application. The specified application is now correctly passed through.
  • Thank you ckarnell for the contribution!

• Fixed MySQL database name reporting when using the database keyword argument

  • Previously, the agent only checked the legacy db keyword argument when parsing the database name from mysql-connector-python connection arguments, so connections made using the modern database keyword argument reported no database name in datastore metrics. The agent now checks database first, falling back to db.

• Fixed warning when accessing LangGraph v1.1+ GraphOutput values

  • Previously, LangGraph v1.1+ instrumentation accessed GraphOutput return values using deprecated item-access syntax, triggering a deprecation warning. The agent now reads GraphOutput.value directly.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds support for Redis commands introduced in Redis v8.10.0 and redis-py v8.1.0. This release also resolves the issue of missing LlmChatCompletion events in LangChain when using MCP tools.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

New features

• Added Redis commands introduced in Redis v8.10.0 and redis-py v8.1.0

  • The Python agent now supports the following recently added commands in Redis:
    • himport_discard_all
    • himport_discard
    • himport_prepare
    • lmovem
    • nrange
    • nrevrange
    • querylabels
    • read
    • sdiffcard
    • sunioncard
    • aliaslist

Bugs

• Fixed missing LlmChatCompletion events in Gemini when using MCP tools

  • Previously, LlmChatCompletionMessages and LlmChatCompletionSummary events were missing in Gemini when using tools in a LangChain agent, if the response format was not the default JSON. This has been resolved.

• Swapped to trace cache ID in LangChain and Agent Framework

  • For context propagation purposes, swap the LangChain and Agent Framework instrumentation to only store the trace cache ID rather than the full trace in a reference attribute. This is to prevent these objects from getting serialized unexpectedly in the case of leaks. The ID is a much smaller object while the trace could contain MBs of data.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds support for AI Monitoring server-side configuration, adds context propagation for Bedrock invocations made through Microsoft Agent Framework, fixes duplicate tool entities for LangChain tools backed by a remote MCP server, reduces race condition risk in the garbage collection data sampler, fixes inconsistent transaction names when using Wagtail, and removes byte encoding from outbound Kombu distributed tracing headers.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

New Features

• Add support for AI Monitoring server-side configuration

  • The Python agent now supports server-side configuration of AI Monitoring. When configured, the server-side ai_monitoring.enabled, ai_monitoring.record_content.enabled, and ai_monitoring.streaming.enabled values take precedence over the local settings.

• Add context propagation for Bedrock invocations through Agent Framework

  • Propagates trace context across thread boundaries created by Microsoft Agent Framework's Bedrock integration. This allows the agent's existing botocore instrumentation to capture LLM events for Bedrock invocations made through Agent Framework. This does not add full instrumentation of Agent Framework itself.

Bugs

• Fix inconsistent transaction names when using Wagtail

  • Previously, transaction names for apps built on Wagtail were inconsistent and could be the same for different page views. This has been fixed.

• Fix duplicate tool entities for LangChain tools backed by a remote MCP server

  • Previously, when a LangChain tool was backed by a remote MCP server that was also separately instrumented, duplicate tool entities appeared in the UI. This has been fixed by gating subcomponent attributes off of LangChain tools when a known remote tool is detected.

• Fix missing User Input field for LangChain and LangGraph MCP tool calls

  • Previously, when LangChain or LangGraph used tools from an MCP server, the User Input field could be missing from AI monitoring data if the input message wasn't the last message in the streamed sequence. This has been fixed, and the agent now correctly extracts the input message in these cases.

• Reduce race condition risk in the garbage collection data sampler

  • Previously, the agent always called gc.get_objects() to sample garbage collector object counts, regardless of configuration. This has been changed so gc.get_objects() is only called when top_object_count_limit is configured above 0, avoiding unnecessary heap iteration for customers who don't track object types collected by the garbage collector.

• Remove byte encoding from outbound Kombu distributed tracing headers

  • Previously, distributed tracing headers added by Kombu instrumentation were byte-encoded, which broke transports that serialize payloads with the standard library json module (for example, SQLAlchemy-backed transports). The byte encoding has been removed so these transports work correctly.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds instrumentation support for Google ADK, adds token counting for multiple AI providers (Anthropic, Google GenAI, OpenAI, and AWS Bedrock), reduces outbound tracestate header size, and fixes a crash in graphql-core instrumentation.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

Improved Features

• Added instrumentation for Google ADK

  • The Python agent now instruments Google ADK (Agent Development Kit) AI agent workflows.
  • Instrumented components include:
    • LlmAgent linked to Gemini or VertexAI backends via google.genai
    • User-defined FunctionTool instances
    • Sub-agent workflows including child agents and AgentTool instances
    • Multi-agent workflows including the Workflow class and deprecated LoopAgent, ParallelAgent, and SequentialAgent classes

• Added token counting for Anthropic, Gemini, OpenAI, and Bedrock

  • The Python agent now reports token counts for all LLM invocations across multiple AI providers:
    • Anthropic SDK
    • Google GenAI SDK
    • OpenAI SDK
    • AWS Bedrock via botocore and aiobotocore

Bugs

• Fixed instrumentation crash for graphql-core v3.2.10

  • Previously, graphql-core v3.2.10 renamed the errors attribute to collected_errors in ExecutionContext, causing the agent to crash. This has been fixed, and the agent now works correctly with graphql-core v3.2.10 and later.

• Removed spanId from outbound tracestate header

  • Previously, the agent included spanId in the outbound tracestate header. Because spanId is already present in the traceparent header, this was unnecessary duplication. The spanId has been removed from tracestate to reduce outbound header size.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds additional model information to LangChain Chain LLM events, and fixes a bug in LangGraph streaming instrumentation.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

Improved Features

• Add additional model information to LangChain Chain LLM events

  • Add request.model and response.model attributes to LLM events produced by LangChain Chains.
  • This information is pulled on a best effort basis and may not always be available.

Bugs

• Fix TypeError in LangGraph streaming instrumentation

  • Previously, when using a subset of streaming APIs in LangGraph, the agent would encounter a TypeError due to the use of weakref.WeakMethod by LangGraph. This has been fixed, and the agent should now work correctly with those LangGraph streaming methods.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent adds support for multiple new Redis commands, fixes a bug in LangGraph instrumentation, and improves handling and logging of harvest limit settings.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

Improved Features

• Add instrumentation for new Redis v8.8.0 commands

  • Add instrumentation for the following new Redis commands from redis-py v8.8.0:
    • Misc commands: INCREX, XNACK
    • Array commands: ARCOUNT, ARDEL, ARDELRANGE, ARGET, ARGETRANGE, ARGREP, ARINFO, ARINSERT, ARLASTITEMS, ARLEN, ARMGET, ARMSET, ARNEXT, AROP, ARRING, ARSCAN, ARSEEK, ARSET

Bugs

• Fix crash in LangGraph instrumentation

  • Previously, when using LangGraph instrumentation, the agent would crash due to an issue with context propagation to jobs in AsyncBackgroundExecutor. This has been fixed, and the agent should now work correctly with LangGraph.

• Improved handling and logging of harvest limit settings

  • The agent now more accurately reports the list of harvest methods and limits to the Environment tab in New Relic One.
  • The agent now has more consistent debug logging on startup for harvest limit settings.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.

Notes

This release of the Python agent fixes a context propagation issue in LangChain and LangGraph, an OpenAI chat completion recording issue, and an issue when using OpenAI(https://pypi.org/project/openai) streaming through LiteLLM as a proxy.

Install the agent using easy_install/pip/distribute via the Python Package Index, using conda via the Conda-Forge Package Index, or download it directly from the New Relic download site.

Bugs

• Fix context propagation in LangChain and LangGraph

  • Adds additional transaction context propagation to LangChain and LangGraph to support a wider range of application structures.

• Fix an OpenAI chat completion recording issue

  • OpenAI reintroduced logic that closes the streaming response even if the consumer hasn't read all data. Previously, the chat completion data was only recorded when reaching the stream's end which resulted in chat completion data not being recorded at the closing of the stream. Logic has now been added to capture the chat completion data even if the stream has not been fully read.

• Fix an issue when using OpenAI streaming through LiteLLM as a proxy

  • Previously when using the OpenAI SDK connected through a LiteLLM proxy server it was possible to lose reporting of LLM response data due to a missing finish reason. This has been fixed, and the response should now always be recorded.

Support statement

We recommend updating to the latest agent version as soon as it's available. If you can't upgrade to the latest version, update your agents to a version no more than 90 days old. Read more about keeping agents up to date.

See the New Relic Python agent EOL policy for information about agent releases and support dates.