Changelog

New Features:

Async MySQL: Added support for AsyncMySQLDb with compatibility for the asyncmy driver. See the docs.
AgentAsJudgeEval: LLM as a judge evaluation system that judges agent outputs against custom criteria with binary (pass/fail) or numeric (1-10) scoring, supporting standalone runs, post-hooks, background execution, and custom evaluator agents. See the docs.
Introduction: Added an Introduction param to Agent and Teams that allows users to set the first “Assistant” message with support for Agent-OS coming soon.
OpenRouter reasoning: Added support for reasoning messages coming from OpenRouter.

Skip database schema creation: You can now set create_schema=False when initializing your database class (e.g. PostgresDb) to skip automatic schema creation. This is useful for environments where the schema needs to be managed externally.
Support get_step_output for nested steps in Parallel groups: Add recursive search to get_step_output() to find steps nested inside Parallel, Condition, Router, Loop, and Steps.
Handle Gemini malformed tool calls: Added logic to automatically detect Gemini’s malformed tool calls, add a guidance message for the model to avoid the error, and retry the failed request.
Support for DeepSeek thinking mode with tools: Added the reasoning_content to the message for thinking mode.
run_id param in run methods: Added a run_id parameter to the run and arun methods in the Agent, Team and Workflow classes, to set the id of the run instead of automatically generating one.
Knowledge Sync Support: Added true sync support for all add_content_ functions instead of wrapping with asyncio

Parallel step does not aggregate inner steps' content: Fixed parallel step in workflow to pass aggregated content to next step and maintain the chained behavior.
Pre-hooks triggering retries: Fixed an error where Agent runs would be unnecessary retried when a InputCheckError was raised with retries=True.

fix: auto db provisioning by @ysolanky in #5619
feat: add AsyncMySQLDb by @harrytran001 in #5425
fix: ensure table auto-creation on async Postgres upserts by @xpander-moriel in #5633
chore: update parallel step to pass aggregated content to next step by @kausmeows in #5636
fix: workflow run not saving to DB with status RUNNING by @kausmeows in #5646
fix: Firecrawl tools limit by @Mustafa-Esoofally in #5639
fix: agent retries by @manuhortet in #5645
fix: get pdf files output by @anuragts in #5316
feat: add create_schema parameter to skip automatic schema creation by @alec-drw in #5649
feat: support get_step_output for nested steps in Parallel groups by @maorkipnis in #5635
chore: add unit tests for the HackerNewsTools toolkit by @edlsh in #5638
[fix] Correct AWS credential validation in Bedrock Claude model by @edlsh in #5653
chore: add a cookbook to show sending traces to different projects in arize by @kausmeows in #5584
feat: add introduction by @ysolanky in #5338
feat: handle Gemini malformed tool calls by @manuhortet in #5652
[fix] support enabling thinking mode with tool calls for deepseek v3.2 by @pengye91 in #5618
feat: async mysql traces by @ysolanky in #5631
feat: support OpenRouter reasoning by @zhiheng-yu in #5641
feat: add run_id parameter to run methods by @alikhtag in #5627
Feat: Implement pure sync knowledge methods by @willemcdejongh in #5642
feat: AgentAsJudge evals [SDK-349] by @harshsinha03 in #5629
chore: Release 2.3.9 by @dirkbrnd in #5651

Full Changelog: v2.3.8...v2.3.9