June 2026 Summaries

Inngest has introduced a new feature called Sessions, aimed at enhancing debugging efficiency in multi-step workflows by allowing users to group related runs under a shared identifier, such as a conversation ID or ticket ID. This feature helps streamline the process of troubleshooting by displaying all related runs together in the Inngest dashboard, providing insights into how many runs occurred, their success or failure, and the functions involved. Sessions are implemented by tagging events with a session key and session ID, allowing for up to five simultaneous groupings per run, which facilitates querying across different dimensions like conversation, tenant, and model. This feature is particularly beneficial for AI agents, as it provides a cohesive view of a user's interaction by linking disparate functions and runs, ultimately enabling session-level scoring to assess the success of conversations. Sessions are available in the public beta of the Inngest TypeScript SDK version 4.7.0 and later, and users can access the feature by updating their SDK and utilizing the meta.sessions object in event sends.

Agent Evals by Inngest introduces a novel approach to evaluating AI agents by focusing on real-world outcomes rather than just the appearance of success. This new system leverages APIs that integrate directly into codebases, allowing for the measurement of outcomes like customer retention and conversion rates, which are not immediately visible after an agent's task completion. It includes features like Experiments, Scoring, and Defer, which enable users to run variant tests, attach meaningful metrics to outcomes, and manage follow-up tasks, respectively. The tool aims to bridge the gap in current observability systems that only evaluate if the code executed correctly without determining if it achieved the desired business results. By incorporating outcome-based scoring directly into the execution layer, Inngest enables more accurate assessments and adjustments in AI models, providing a more reliable means of determining effectiveness. This approach not only enhances model observability but also aligns technical performance with business objectives, ensuring that agents contribute positively to the bottom line.

Inngest has introduced a new scoring API that allows developers to attach verdicts to any run or step in their applications, enhancing the ability to measure outcomes rather than just completion. This feature, step.score(), integrates directly into existing codebases, enabling real-time scoring of every run without the need for external observability tools or manual processes. By embedding the scoring mechanism within the execution layer, Inngest allows users to evaluate real data flows, ensuring that the scoring reflects actual outcomes rather than proxy indicators. The API supports deferred scoring, which can wait for external events to occur before assigning a score, and integrates with the new group.experiment() feature to evaluate different strategies in live production environments. This approach contrasts with traditional evaluation tools by eliminating the need to reconcile data across separate systems and reducing the cost and complexity associated with sampling and storing data externally. Scoring is available in beta with version 4.8 of the TS SDK, offering a seamless way for developers to implement outcome-based evaluations directly within their workflows.

Lauren Craigie's blog article introduces a new SDK primitive called group.experiment() by Inngest, designed to enable safe testing of changes to production code by comparing different variants on live traffic using execution data. This tool allows developers to measure and analyze the performance of various implementations, such as workflow steps or data transformations, by leveraging execution data like timing, status, and retries, which are recorded for every run. Unlike traditional experimentation tools that rely on input and output data, group.experiment() is integrated directly into the execution layer, allowing for a more comprehensive comparison of variants based on real production metrics such as speed, reliability, and cost-effectiveness. This approach eliminates the need for rewrites or pre-wired data pipelines, offering a more efficient and accurate method for determining the best variant. The functionality supports multiple selection strategies, including weighted random splits, bucket-based assignments, custom logic, and fixed selections, providing flexibility in experimentation and deployment scenarios. The tool is available in the Inngest TypeScript SDK 4.8.0 and later, encouraging users to upgrade to take advantage of its features.

The blog article by Aaron Harper discusses the implementation of Node.js worker threads in the production environment of Inngest Connect to maintain connection health independent from potentially blocking user code. The transition involved moving critical components like the WebSocket connection, heartbeat timer, reconnect logic, and lease extension logic into a separate event loop handled by a worker thread, while user code remained on the main thread. This separation required careful design considerations, such as defining clear API boundaries with postMessage, converting callbacks into a request-response protocol, and ensuring all data exchanged between threads is structured-clone safe. Additionally, the article highlights the importance of creating robust systems around the worker threads, including message protocols, logger proxying, crash recovery, and graceful shutdown processes. This approach ensures that critical timers and connection management tasks continue to function even if the main thread is busy, ultimately enhancing the reliability and stability of the Connect service.

The blog article explores the concept of agent loop architecture, emphasizing its importance in AI systems by detailing its three fundamental layers: loops, skills, and orchestrators. It underscores the significance of durable orchestration, which ensures that agent loops can withstand failures, process restarts, and infrastructure changes without losing progress. The text explains how loops act as decision-makers triggered by schedules, skills represent durable workflows that enhance loop capabilities, and orchestrators manage the execution, retries, and deployment of these functions. The discussion highlights the necessity of observability and step-level checkpointing to maintain efficiency and reduce costs. It also presents a vision where AI agents are empowered to autonomously evolve by writing and deploying new functions, thereby enhancing their operational capabilities. The article concludes by advocating for the integration of durable execution primitives, suggesting that this approach is fundamental to developing resilient and adaptable AI systems.

Insights, a feature within the Inngest dashboard, has expanded its capabilities to allow users to query not only event data but also runs, steps, step attempts, and trace spans, providing a deeper understanding of function executions and debugging processes. Built on ClickHouse for fast querying, Insights enables teams to access detailed execution data directly without needing to export it, enhancing observability and troubleshooting. The recent update allows users to query inside function runs and access SQL JSON fields, making it possible to track specific functions, errors, and performance issues. Additionally, Insights AI has been integrated into the sidebar to facilitate query creation through plain English descriptions, while Schema Explorer provides an updated view of data payload structures. With these enhancements, Insights offers a comprehensive tool for analyzing workflow triggers and execution details, supported by the Inngest Realtime API and durable execution capabilities.

Inngest has introduced new command-line interface (CLI) commands and a version 2 REST API to enhance the debugging and execution tracking capabilities of its function runs, accessible directly from the terminal, CI pipelines, or coding agents. These new CLI commands, organized under the `inngest-cli api`, allow users to programmatically access function runs, traces, and invocations with a new API key type for authentication, providing real-time execution data from local dev servers or Inngest Cloud. The CLI outputs compact JSON, which can be integrated into debugging workflows for seamless error tracking and resolution, allowing developers to pull run summaries, step traces, and invoke functions locally. Additionally, these CLI commands are built on the Inngest v2 REST API, which is publicly documented, allowing developers to interact directly with the API for custom script and tool integrations. API keys are now used for programmatic access, and the documentation provides setup guidance for both the CLI and coding agents, promising further command enhancements to improve user experience.

JavaScript developers face an overwhelming variety of schema libraries, such as Zod, Valibot, ArkType, and others, which complicates the decision-making process for library authors who need to support user-defined schemas. The introduction of Standard Schema offers a solution by providing a single compatibility interface that many popular schema libraries already implement, allowing developers to support multiple libraries without the need for specific adapters. This approach does not introduce a new schema language or require learning new syntax; instead, it allows existing schema libraries to validate unknown input and recover input/output types. Standard Schema focuses on runtime validation and static type inference, leaving error formatting, metadata, and other aspects to be handled by the specific libraries. It simplifies the integration process within the TypeScript SDK environment by standardizing validation interactions across different schema libraries, thereby reducing the friction for users and library maintainers alike.

defer() is a new API introduced by Inngest, designed to enhance orchestration tools by allowing for typed, durable follow-up functions to be launched from within a parent run. This innovation addresses the challenge of disconnected follow-up processes by providing a seamless mechanism to pass necessary data, pause execution for extended periods, and link deferred runs to their parent in the user interface. The API enables efficient post-completion background work, overcoming traditional limitations of job queues, workflow engines, and evaluation platforms, which often require separate handlers and manual data alignment. defer() operates as a fire-and-forget mechanism, buffering operations until the parent run finalizes, thus preventing any interference with the parent process. It offers comprehensive functionality, including step execution and event waiting, while ensuring data validation through typed payloads. Currently available in beta within the Inngest TypeScript SDK, defer() is poised to streamline workflows by integrating closely with the rest of Inngest's API, promising future enhancements like batching and encryption support.

Inngest has launched official tooling for coding agents like Claude Code, Codex, and Cursor, enabling developers to create more durable infrastructure directly within their existing coding environments. This integration allows for the seamless execution and testing of Inngest code, including setup, events, durable functions, and flow control, without needing to switch tools. The plugins provide real-time updates and maintain the latest guidance, ensuring that developers have access to current step patterns and retry behaviors necessary for production-grade systems. The collaboration between Inngest and these agents is designed to enhance the durability of background processes, webhooks, and AI workflows, with the open-source project actively seeking community feedback to evolve and expand its capabilities.