Google AI Studio adds Firebase as Cursor releases Composer 2

Google expanded the build features in Google AI Studio in mid-March 2026. A new coding agent can make multi-step code changes and connect apps to external services. It uses components from Google Antigravity.¹ With the Firebase integration, AI Studio can identify when a prompt calls for a database or login and set up the required services with user approval.²¹

The changes let developers add stored data and user accounts to apps created from prompts.²¹ Google describes the update as a step toward production use.¹ Companies still need to determine whether a generated app meets their requirements. That includes reviewing its code, access permissions, and connections to external services before deployment.

Cursor introduced Composer 2 on March 19 as a coding model for difficult, multi-step development tasks. It is available inside Cursor, where a faster variant with what the company says is the same intelligence is the default. Cursor is also positioning the model as a lower-cost alternative to other leading coding systems.¹²

A technical report published March 27 gives Composer 2 a score of 61.3 on CursorBench, up 37% from Composer 1.5. That is Cursor’s own measurement, not an independent comparison.³ Composer 2 also builds on an open-source version of Moonshot AI’s Kimi. Cursor said the base accounted for about a quarter of the final model’s compute, with its own training accounting for the rest. How much each stage contributes to performance remains unclear.⁴

Anthropic has introduced Claude Code Channels as a research preview. Developers can send messages from Telegram or Discord to a running Claude Code session and receive replies in the same chat.¹² Official plugins are available for both messaging services, and the connection requires an active session.¹³ That lets developers follow a coding task when they are away from the terminal.¹

The connection adds another way to reach Claude Code, but it does not replace the running development environment or its setup.¹³ Available descriptions say the preview requires a claude.ai login; access using only an API key is not supported.¹⁴ For teams, the practical question is which messaging accounts may send instructions to a session and which development environment that session can access.¹³

OpenAI plans to bring ChatGPT, Codex, and its Atlas browser into a single desktop application, according to The Wall Street Journal and Bloomberg.¹² Both outlets reported the plan on March 19, 2026, without giving a launch date.¹² The proposed consolidation is not a released application.¹²

A shared interface could simplify access to OpenAI’s currently separate chat, coding, and web tools.¹² The reports also describe a desire to focus resources more closely on productivity uses.¹² For companies, the distinction between a plan and an available product matters: the reporting does not establish which functions will work together when the application launches. Existing workflows therefore cannot yet be assessed against a defined replacement.¹²

MiniMax introduced M2.7 on March 18, 2026, and made it available through MiniMax Agent and the MiniMax API Platform that day. The company calls it its first model to participate deeply in its own development. MiniMax reports a score of 56.22% on SWE-Pro.¹²

MiniMax positions M2.7 as a successor to M2.5, emphasizing software engineering and agent tasks. Its published materials do not explain the training process in enough detail to establish how that participation worked. In particular, they leave open how much of the improvement came from model-written training code, feedback loops, or human direction. The reported performance figures are company claims.¹²

Baidu introduced Qianfan-OCR on March 18, 2026. The 4-billion-parameter document model is available with open weights on Hugging Face and through the Baidu Qianfan platform.¹²³ It aims to combine text recognition, layout analysis, and extraction of tables, formulas, and key information in one model.²⁴ It can convert document images directly to Markdown, with an optional “Layout-as-Thought” step to assess positions and reading order first.⁴

Baidu is targeting workflows that currently connect several document-processing stages.² The company says a single vLLM instance can handle processing without multistage orchestration.² Its published performance figures come from Baidu and the model materials; the supplied sources do not establish independent verification.²⁵ Companies will need to test whether the model simplifies their workflows using their own documents and operating requirements.

Anthropic published findings in March 2026 from interviews with Claude users in 159 countries. A specially configured Claude “Interviewer” held open-ended conversations about how participants use AI, what they want from it, and what concerns them. The conversations took place in December 2025, and Anthropic also used Claude to classify the responses. Participants most often wanted help with productivity and doing better professional work. Their leading concern was unreliable or incorrect output, ahead of job displacement.¹²

The findings offer a view of the priorities and concerns of active users. Responses varied by region: participants in sub-Saharan Africa and parts of Asia were generally more optimistic than those in Western Europe and North America.³ The study is not a representative survey of the global population. It drew on Claude users who chose to participate, so its findings cannot automatically be generalized to everyone using AI.¹

Jeff Bezos is in early talks about a fund targeting up to $100 billion, according to reports published March 19. The proposed fund would buy manufacturers and use AI to advance automation at their businesses.¹² No completed fundraising or fund launch has been announced, and the discussions remain preliminary.³²

The fund would be separate from Project Prometheus, the Bezos-backed AI company working on engineering and manufacturing applications. Its reported areas of interest include computers, cars, and spacecraft.³ The plan would pair the acquisition of industrial businesses with the deployment of AI inside them, rather than funding AI development alone. Whether investors will commit the proposed amount—and which businesses might be acquired—remains unclear.¹³

Google Labs announced a broad Stitch update on March 18, 2026. Its new canvas lets users work on UI designs and interact with the workspace by voice; the voice capability is initially available in Preview. Stitch can provide design critiques as users work and make live changes. Voice is part of the wider product update, not a separate new tool.¹²

The update also includes an infinite canvas, a design agent that tracks progress, and prototype generation.¹ Google is moving more of the design process into an ongoing conversation with the tool. For teams, the practical question is whether its suggestions can be directed reliably when requirements become more complex. The announcement does not provide comparative quality results for the new features.¹

Princeton University researchers introduced OpenClaw-RL in March 2026. The approach is designed to train AI agents continuously from interactions while they keep handling requests.¹² It runs agent use and training asynchronously, so new feedback can inform subsequent updates without stopping the agent for each training step.¹

The work addresses a practical problem in agent operations: Feedback arrives throughout real tasks, but it does not automatically become a useful training signal.¹ OpenClaw-RL aims to extract such signals from user replies and changes in the agent’s working environment.¹ That could make it easier to adapt agents to recurring tasks. It remains unclear how reliably the method learns over long periods or performs across different deployment settings.