BLOG.JETBRAINS.COM
Building Better Agents: Whats New in Koog 0.3.0
Weve just released Koog 0.3.0, which comes with many updates that make building, running, and managing intelligent agents easier.This version focuses on durability, speed, observability, and smoother integration with real-world systems. If youve been exploring how to develop your own intelligent agent that can handle complex workflows, the new and updated Koog could be just what youre looking for. Find it now on GitHub.Heres a breakdown of whats new in v0.3.0:Learn moreAgents that dont forgetIn Koog 0.3.0, agents can now remember what they were doing. You can save and reload their state from a local disk, S3 buckets, or a database. This powerful feature enables server-side developers to build fault-tolerant agentic solutions by restoring the agents entire state machine at the exact point of strategy execution instead of just recovering the message history.This means you can shut down machines without losing progress. Theres also a new checkpoint feature, so agents can roll back to any earlier state if needed.Smarter storage for better retrievalKoog now supports persistent vector storage for docs. Whether youre working with local files or connecting to a vector database, you can use Koog to build retrieval-augmented generation (RAG) pipelines.The setup is modular, allowing you to add support for different formats, ranking methods, and backends.Built-In observability and moderation for agentsWeve added native OpenTelemetry support. This means you can now trace, log, and measure your agents using the same tools youd use elsewhere in your system, enabling you to spot bottlenecks, track behaviour, and keep things running smoothly.And if youre deploying in regulated environments, theres also built-in support for moderation. Agents can now check their own outputs to make sure theyre appropriate and safe. This is especially useful when trust and compliance are a priority.Running tasks in parallelSometimes your agent graph includes steps that can run independently. Koog now supports parallel execution of those nodes, following a familiar MapReduce-style API.You can launch several branches simultaneously, transform their results asynchronously, and collect everything at the end. This approach is useful when you want to try multiple strategies at once or speed up multi-step workflows.A better fit for your stackKoog now works better with Spring. If youre using Spring Boot, youll get ready-made beans and auto-configured LLM clients out of the box.The release also adds support for ReAct-style agent thinking. Agents can now follow step-by-step reasoning paths, switching between thoughts and actions.Youll also find updates that help agents handle uncertainty better. The new Retry component makes them more resilient, and multiple-choice reasoning gives them new ways to explore options and respond to user preferences.A few more things to get excited aboutThis version brings a handful of other improvements that are worth highlighting:Agents can now receive image input when running with Ollama-backed models.Full WebAssembly (WASM) support means you can now deploy Koog agents to the browser.Native support for Amazon Bedrock has been added.Fact-retrieval history compression has been improved to help agents focus on the correct data without getting confused by unrelated context.Agents can now work with arbitrary input and output types, providing greater flexibility in the execution flow.Weve also fixed several bugs to make the overall UX smoother.Koog 0.3.0 is a step toward making agents more robust, portable, and valuable.Many of these updates were implemented on the basis of community feedback, so thanks to everyone whos tried it out, opened issues, or shared ideas. If you havent yet, nows a good time to try Koog and see what you can build.Your contributions make the differenceWed like to take this opportunity to extend a huge thank-you to the entire community for contributing to the development of Koog through your feedback, issue reports, and pull requests!Heres a list of this releases top contributors:Nathan Fallet Azure OpenAI and other general improvementsDidier Villevalois Ollama updatesJason Pearson AWS Bedrock improvementsDenys Kurylenko MCP enhancements
0 Commentaires 0 Parts 89 Vue