Technology Radar Vol 34
Languages & Frameworks
Adopt
-
98. Apache Iceberg
Apache Iceberg is an open table format for large-scale analytical datasets that defines how data files, metadata and schemas are organized on storage systems such as S3. Having evolved significantly in recent years, it has become a foundational building block for technology-agnostic lakehouse architectures.
Iceberg is now supported by all major data platform providers — including AWS (Athena, EMR, Redshift), Snowflake, Databricks and Google BigQuery — making it a strong option for avoiding vendor lock-in. What distinguishes Iceberg from other open table formats is its openness across features and governance, unlike alternatives whose capabilities are limited or controlled by a single vendor.
From a reliability perspective, Iceberg's snapshot-based design provides serializable isolation, safe concurrent writes through optimistic concurrency and version history with rollback. These capabilities deliver strong correctness guarantees while avoiding performance bottlenecks.
While Apache Spark remains the most common engine used with Iceberg, it’s also well supported by Trino, Flink, DuckDB and others, making it suitable for a wide range of use cases, from enterprise data platforms to lightweight local analytics. Across many of our teams, Iceberg has earned strong trust as a stable, open data format; we recommend it as a default choice for organizations building modern data platforms.
-
99. Declarative Automation Bundles
Declarative Automation Bundles (formerly known as Databricks Asset Bundles) has evolved into a primary tool for bringing software engineering and CI/CD practices to the Databricks ecosystem. The tool has matured significantly and now enables our teams to manage most platform resources as code, including clusters, ETL pipelines, jobs, machine-learning models and dashboards. Using the databricks
bundle plancommand, teams can preview changes and apply repeatable deployment practices to Databricks artifacts, similar to how infrastructure is managed with tools such as Terraform. Treating traditionally mutable assets, such as dashboards and ML pipelines, as code allows them to be version-controlled, tested and deployed with the same rigor as traditional microservices. Based on our experience in production environments, Declarative Automation Bundles has become a reliable approach for managing data and ML workflows on Databricks. Teams working extensively in the Databricks ecosystem should consider adopting it to standardize infrastructure management practices. -
100. React JS
React has been our default choice for JavaScript UI development since 2016, but with the stable release of React Compiler coming last October (as part of React 19), it's worth revisiting. There are a number of reasons this feature is notable. By handling memoization at build time, for example, it makes manual
useMemoanduseCallbacklargely unnecessary, though the team recommends keeping them as escape hatches when precise control over effect dependencies is required. Battle-tested at Meta and supported by Expo SDK 54, Vite and Next.js, the compiler also removes a category of performance boilerplate that has long been a cost of working at scale with React. React 19 also introduces Actions and hooks such asuseActionStateanduseOptimistic, which simplify form handling and data mutations without relying on external libraries. The launch of the React Foundation under the Linux Foundation in 2025 — with Amazon, Expo, Callstack, Microsoft, Software Mansion and Vercel joining Meta — further strengthens the library's long-term stability and addresses a concern cautious teams have historically cited when considering adoption. -
101. React Native
React Native has moved to Adopt as our default choice for cross-platform mobile development. While previously in Trial, the rollout of the New Architecture — specifically
JSIandFabric— has addressed long-standing concerns regarding bridge bottlenecks and initialization speed. Our teams have observed significant performance gains in complex UI transitions and data-intensive workloads. By moving away from the asynchronous bridge, React Native now delivers responsiveness that rivals native implementations while maintaining a single codebase.We’ve used it successfully across multiple production projects and found the ecosystem around Expo and React to be mature and stable. While state management still requires careful planning, the productivity benefits of
fastrefreshworkflows and shared skill sets outweigh these costs. For most hybrid mobile use cases, React Native is now our primary recommendation for teams seeking performance, consistency and speed. -
102. Svelte
Svelte is a JavaScript UI framework that compiles components into optimized JavaScript at build time rather than relying on a large browser-side runtime or virtual DOM. Since we last featured it in Trial, we've seen more teams use it successfully in production. SvelteKit has also made it a more robust choice for SSR and full-stack web applications, increasing our confidence in moving it to Adopt.
In our experience, the original reasons to choose Svelte still hold: it produces small bundles, delivers strong run-time performance and offers a simpler component model. Newer capabilities in Svelte 5, such as runes and snippets, make reactivity and UI composition more explicit and flexible. Compared with heavier front-end frameworks, Svelte provides a cleaner development experience with less code. Feedback from teams increasingly suggests it’s a credible alternative to React or Vue rather than a niche option. Teams should still consider ecosystem familiarity, hiring and platform fit, but we now recommend Svelte as a sensible default for building modern web applications where performance and delivery simplicity matter.
-
103. Typer
Typer is a Python library for building command-line interfaces from standard type-annotated functions, providing automatic help text, shell completion and a clear path from small scripts to larger CLI applications. We're seeing increased relevance as teams turn internal tooling, automation and AI-adjacent developer workflows into first-class CLIs. In our experience, Typer is easy to introduce on real projects, and teams appreciate how quickly it produces clear, readable commands. Its strengths include type-hint-driven APIs, automatic help and completion and a low-friction path from simple scripts to multi-command CLIs. However, it’s a Python-specific solution and may not be the best fit when highly customized CLI behavior or cross-language consistency is required. We recommend Typer for teams building CLIs for delivery, operations and developer experience workflows.
Trial
-
104. Agent Development Kit (ADK)
Agent Development Kit (ADK) is Google’s framework for building and operating AI agents with software-engineering-oriented abstractions for orchestration, tools, evaluation and deployment. Its ecosystem and operational capabilities have matured significantly since we included it in Assess, with active multi-language development and stronger observability and run-time features. Vendor-native agent frameworks are now a crowded field, with Microsoft Agent Framework, Amazon Bedrock AgentCore, the OpenAI Agents SDK and the Claude Agent SDK advancing competing options. Open-source alternatives such as LangGraph and CrewAI remain strong choices, especially where teams prioritize framework portability and broader ecosystems. ADK remains pre-GA in parts, with occasional rough edges and upgrade friction, but we’ve seen more projects using it successfully, particularly those already invested in Google’s platform.
-
105. DeepEval
DeepEval is an open-source, Python-based framework for assessing LLM performance. It can be used to evaluate retrieval-augmented generation (RAG) systems and applications built with frameworks such as LlamaIndex or LangChain, as well as to baseline and benchmark models. DeepEval goes beyond simple word-matching metrics, assessing accuracy, relevance and consistency to provide more reliable evaluation in real-world scenarios. It includes capabilities such as hallucination detection, answer relevance scoring and hyperparameter optimization. One feature our teams have found particularly helpful is that it allows teams to define custom, use-case–specific metrics.
Recently, DeepEval has expanded to support complex agentic workflows and multi-turn conversational systems. Beyond evaluating final outputs, it provides built-in metrics for tool correctness, step efficiency and task completion, including evaluation of interactions with MCP servers. It also introduces conversation simulation to automatically generate test cases and stress-test multi-turn applications at scale.
-
106. Docling
Docling is an open-source Python and TypeScript library for converting unstructured documents into clean, machine-readable outputs. Using a computer vision–based approach to layout and semantic understanding, it processes complex inputs — including PDFs and scanned documents — into structured formats such as JSON and Markdown. That makes it a strong fit for retrieval-augmented generation (RAG) pipelines and for producing structured outputs from LLMs, in contrast to vision-first retrieval approaches such as ColPali.
Docling provides an open-source, self-hostable alternative to proprietary cloud-managed services such as Azure Document Intelligence, Amazon Textract and Google Document AI, while integrating well with frameworks such as LangGraph. In our experience, it performs well in production-scale extraction workloads across digital and scanned PDFs, including very large files containing text, tables and images. It delivers a strong quality-to-cost balance for downstream agentic RAG workflows. Based on these results, we’re moving Docling to Trial.
-
107. LangExtract
LangExtract is a Python library that uses LLMs to extract structured information from unstructured text based on user-defined instructions, with precise source grounding that links each extracted entity to its location in the original document. It processes domain-specific materials such as clinical notes and reports. A key strength is source traceability, which ensures each extracted data point can be traced back to its source. The extracted entities can be exported as a JSONL file, a standard format for language model data, and visualized through an interactive HTML interface for contextual review. Teams considering structured output from LLMs for document processing should evaluate LangExtract alongside schema-enforcement approaches such as Pydantic AI. LangExtract is better suited to long-form, unstructured source material, while Pydantic AI excels at constraining output formats for shorter, more predictable inputs.
-
108. LangGraph
Since the previous Radar, we’ve observed that the LangGraph architecture — which treats every multi-agent system as stateful graphs with a global shared state — is not always the best approach for building agentic systems. We’ve also seen an alternative approach, used in frameworks such as Pydantic AI, that also works well.
Instead of starting with a rigid graph and a massive shared state, this approach favors simple agents communicating through code execution, with graph structures added later when needed. It often results in leaner and more effective systems for many use cases. Because each agent only has access to the state it needs, reasoning, testing and debugging become easier. As a result, we’ve moved LangGraph out of Adopt. While it remains a powerful tool, we no longer see it as the default choice for building every agentic system.
-
109. LiteLLM
LiteLLM started as a thin abstraction layer over multiple LLM providers but has since expanded into a full-fledged AI gateway. Beyond simplifying API integration, it addresses common cross-cutting concerns in GenAI systems such as retries and failover, load balancing across providers and cost tracking with budget controls.
Our teams are increasingly adopting LiteLLM as a sensible default for AI-powered applications. The gateway provides a consistent place to address governance concerns, including request tracing, access control, API key management and edge-level guardrails such as content filtering and data redaction or masking. However, teams that rely on differentiating provider features will find these often require provider-specific parameters, reintroducing the coupling the gateway is meant to eliminate. It’s also worth noting that
drop_paramsmode silently discards unsupported parameters, meaning capabilities may be lost across routing decisions without visibility. LiteLLM remains a pragmatic choice for operational control, but teams should understand that leaning into provider-specific capabilities means maintaining both a gateway dependency and provider-coupled code. -
110. Modern.js
Modern.js is a React meta-framework from ByteDance that we’re placing in Trial for teams with micro-frontend requirements built on Module Federation. The trigger is practical:
nextjs-mfis heading toward end-of-life. The Pages Router will receive only small backported fixes, no new development is planned, and CI tests are expected to be removed by mid-to-late 2026. With Next.js lacking official Module Federation support and the community plugin being phased out, the Module Federation core team now recommends Modern.js as the primary supported framework for federation-based architectures. The@module-federation/modern-js-v3plugin provides automatic build wiring out of the box, with streaming SSR and Bridge APIs available as separate capabilities. However, combining them has limitations:@module-federation/bridge-reactis not yet compatible with Node environments, making Bridge unusable in SSR scenarios.Our early experience is positive, and the migration path is well defined for teams already using Module Federation. The ecosystem outside ByteDance is still maturing, so teams should plan for thinner documentation and closer engagement with upstream. This remains a Trial recommendation, because investment is justified for Module Federation use cases where no better-supported alternative currently exists.
Assess
-
111. Agent Lightning
Agent Lightning is an agent optimization and training framework that enables automatic prompt optimization, supervised fine-tuning and agentic reinforcement learning. Most agent frameworks focus on building agents, not improving them over time. Agent Lightning addresses this by enabling teams to continuously improve existing agents without changing their underlying implementation, as it supports frameworks such as AutoGen and CrewAI. It achieves this through an approach called Training-Agent Disaggregation, which introduces a layer between the training and agent frameworks. This layer consists of two core components: the Lightning Server and the Lightning Client. The Lightning Server manages the training process and exposes an API for updated models, while the Lightning Client acts as a runtime that collects traces and sends them back to the server to support training. We recommend teams with established agent deployments explore Agent Lightning as a way to continuously improve agent performance.
-
112. GitHub Spec Kit
Spec-driven development featured prominently in our discussion this cycle. Two broad camps are emerging: teams that rely on the continually improving capabilities of coding agents with minimal structure, and those that favor defined workflows and detailed specifications.
Several of our teams are experimenting with spec-driven practices using GitHub Spec Kit, mostly in brownfield environments. A key concept in Spec Kit is the constitution, a foundational rulebook that aligns the software development lifecycle. In practice, teams reported that a useful constitution typically captures project scope, domain context, technology versions, coding standards and repository structure (for example, hexagonal architecture or layered modules). This shared context helps agents operate within the intended architectural boundaries.
Teams also encountered challenges such as instruction bloat, where continually adding project context led to a growing agent instruction set and eventually context rot. One team addressed this by extracting reusable guidance into skills, keeping agent instructions lean and loading detailed context only when needed.
In brownfield systems, much rework stems from unclear intent, hidden assumptions and late discovery of constraints. One team adopted a lifecycle of spec → plan → tasks → coding → review, which helped surface issues earlier. Over time they also moved repeatable context into files such as
.github/prompts/speckit.<command>.prompt.md, making prompts shorter and agent behavior more consistent. Teams did report rough edges, including unnecessary defensive checks and overly verbose markdown outputs that increased cognitive load. Customizing Spec Kit templates and instructions — for example limiting the number of generated markdown files and reducing console verbosity — helped address some of these issues. Ultimately, experienced engineers, particularly those with strong clean coding and architectural practices, tend to extract the most value from spec-driven workflows. -
113. Mastra
Mastra is an open-source, TypeScript-native framework for building AI applications and agents. It offers a graph-based workflow engine, unified access to a variety of LLM providers, human-in-the-loop suspension and resumption, as well as RAG and memory primitives. It also includes MCP server authoring and built-in tooling for evaluation and observability, supported by clear developer documentation. Mastra provides an alternative to Python-heavy stacks, allowing teams to build feature-rich AI capabilities directly within existing web ecosystems such as Node.js or Next.js. It's worth evaluating for teams already invested in the TypeScript ecosystem who want to avoid switching to Python for their AI layer.
-
114. Pipecat
Pipecat is an open-source framework for building real-time voice and multimodal agents with a modular pipeline model for STT, LLM, TTS and transport orchestration. It’s attracting strong interest because teams can iterate quickly on conversational behavior and swap providers with relatively low friction.
Compared to LiveKit Agents, Pipecat offers greater framework flexibility but a less integrated production path, especially for self-hosted deployment, transport reliability and low-latency turn handling at scale. We’ve placed Pipecat in Assess because it provides a strong engineering-facing foundation, but significant platform engineering work is required before it can be relied on for business-critical production workloads.
-
115. Superpowers
With the growing use of coding agents, there is no single prescribed workflow for every team. Instead, teams are evolving bespoke workflows based on their context and constraints. Our teams across the globe are actively defining and experimenting with such AI-assisted coding workflows. Superpowers is one such workflow built from composable skills. It wraps a coding agent in a structured workflow of skills that encourage brainstorming before coding, detailed planning before implementation, test-driven development with enforced red-green-refactor cycles, systematic root-cause-first debugging and post-implementation code review. Superpowers is distributed as a plugin via the Claude Code plugin marketplace as well as the Cursor plugin marketplace.
-
116. TanStack Start
TanStack Start is a full-stack framework built on TanStack Router for React and Solid. It's comparable to Next.js, supporting SSR, caching and many of the same features. TanStack Start provides end-to-end compile-time safety across server functions, loaders and routing, reducing the risk of broken links or mismatched data shapes on the frontend.
The framework favors explicit configuration over convention, making the experience closer to working with plain React. You can also progressively add SSR capabilities as needed. Compared to Next.js, which has more opinionated defaults that can lead to unexpected behavior if teams are unfamiliar with its inner workings, TanStack Start is more explicit and predictable. The TanStack ecosystem has also matured significantly, offering a strong set of tools for building modern web applications.
-
117. TOON (Token-Oriented Object Notation)
TOON (Token-Oriented Object Notation) is a human-readable encoding for JSON data designed to reduce token usage when structured data is passed to LLMs. It allows teams to retain JSON in existing systems and transform it only at the point of interaction with the model. This matters because token cost, latency and context-window constraints are becoming real design considerations in RAG pipelines, agent workflows and other AI-heavy applications. Raw JSON often spends tokens on repeated keys and structural overhead rather than useful content.
In our early evaluation, TOON is an interesting last-mile optimization for prompt inputs, particularly for large, regular datasets where a more schema-aware format can be both more efficient and easier for models to process than JSON. It’s not a replacement for JSON in APIs, databases or model outputs, and is often the wrong choice for deeply nested or non-uniform structures, semi-uniform arrays or flat tabular data where CSV is more compact. It may also be less suitable for latency-critical paths where compact JSON performs well. For these reasons, we think TOON is worth assessing for teams building LLM applications where structured input size is a meaningful cost or quality concern. Teams should benchmark it against JSON or CSV using their own data and model stack.
-
118. Unsloth
Unsloth is an open-source framework for LLM fine-tuning and reinforcement learning that focuses on making training significantly faster and more memory efficient. Fine-tuning LLMs involves billions of matrix multiplications, workloads that benefit from GPU acceleration. Unsloth optimizes these operations by translating them into highly efficient custom kernels for NVIDIA GPUs, dramatically reducing cost and memory usage. This makes it possible to fine-tune models on consumer GPUs such as T4 and above rather than requiring expensive H100 clusters. Unsloth supports LoRA, full fine-tuning, multi-GPU training and long-context fine-tuning (up to 500K tokens) for popular models including Llama, Mistral, DeepSeek-R1, Qwen and Gemma. As domain-specific AI applications increasingly rely on fine-tuning, Unsloth significantly lowers the barrier.
Caution
Unable to find something you expected to see?
Each edition of the Radar features blips reflecting what we came across during the previous six months. We might have covered what you are looking for on a previous Radar already. We sometimes cull things just because there are too many to talk about. A blip might also be missing because the Radar reflects our experience, it is not based on a comprehensive market analysis.
