Skip to main content
Back

Replit launches Agent 3 with 200-min autonomous runtime

OpenAI adds MCP in ChatGPT, Anthropic ships web fetch + fixes Claude quality, Stability AI's audio model, ElevenLabs tes

September 11, 2025 4 min read

Hey

Welcome to AlphaSignal, the most read source of news by AI engineers and researchers.


Every day, we identify and summarize the top 1% of news, papers, models, and repos, so you’re always up to date.


Here’s today’s roundup:

SUMMARY

Read time: 4 min 25 sec

Top News

▪ Replit introduces Agent 3, its most autonomous coding agent yet

Nylas

▪ Learn how AI meeting tools drive adoption and retention in 2025

Trending Signals

▪ AlphaSignal PRO is coming: help us decide what goes into it

▪ OpenAI enables MCP connectors in ChatGPT developer mode

▪ Anthropic adds web fetch in Claude API to analyze webpage content

▪ Stability AI debuts audio model for controllable sound production at scale

▪ ElevenLabs releases voice editing tool in alpha for agent voice customization

▪ Anthropic fixes Claude quality problems, clarifies no deliberate reduction

Trending Papers

▪ Google introduces AI system to automate code generation for science

▪ OpenAI and Anthropic jointly study safety of Claude and GPT models

▪ Google study finds where embeddings fail

Top Read

▪ Mira Murati’s lab finds cause of inconsistent LLM outputs

AlphaSignal PRO is coming


We’re building a new version to help you learn, and implement the latest AI breakthroughs.


Take 2 minutes to tell us which features matter most, your feedback will directly decide what goes into AlphaSignal Pro.

Give feedback

TOP_NEWS

Replit debuts Agent 3: 200-minute long runtime, agent creation, and browser-native application testing

5,294 Likes

alpha_signal_image_2

Replit introduced Agent 3, its upgraded autonomous coding agent, giving users the ability to generate other agents and automations.

The system tests apps directly in the browser, auto-fixes issues, and runs 10× more autonomously than Agent V2.


Key Features

  • In-Browser App Testing: Agent runs UI tests, checks APIs/forms, and auto-fixes issues.

  • Max Autonomy (Beta): Extended sessions (~200 min) with task management and self-supervision.

  • Agents & Automations: Build bots and workflows for Slack, Telegram, Notion, Linear, etc.

  • Flexible App Modes: Choose frontend-only prototyping or full-stack builds.

How It Works

  • Agent iterates in a build → run in browser → validate → repair loop. It decides when enough has changed to re-test rather than testing after every prompt.

  • The testing system runs inside Replit’s workspace preview and reports failures for automatic repair.

Performance

  • Replit reports the testing system is 3× faster and 10× more cost-effective than generic Computer-Use approaches.

  • Autonomous runtime: up to ~200 minutes per session for long tasks.

Example Usecases

  • Slack bot that answers codebase questions via GitHub integration.

  • Notion Q&A agent

  • Linear daily roll-up email

  • meeting-prep agent that searches the web and files notes to Drive

Availability

  • Available now to all free and paid users in Replit

  • App Testing is enabled by default (can be turned off).

  • Max Autonomy is labeled beta.

Note

Replit announced Agent 3 alongside a $250M fundraise at a $3B valuation.

Try now
alpha_signal_image_3

Unlock 2025 Insights: 68% Will Switch for Better Meeting Recording

Nylas shows why meeting recording isn’t optional, 68% of SaaS users would switch platforms for better meeting features. Explore findings from 1,000 surveyed and see what’s shaping the future of SaaS platforms.

  • 85% report more efficient meeting follow-ups with recording tools

  • 89% of teams have AI meeting features on their roadmap

  • Integrated solutions increase user retention and reduce context switching

  • 41% already build AI meeting features today

  • Privacy and compliance are top adoption drivers in regulated sectors

Read the latest Nylas report to get direct insights and proven trends.

Read the report

partner with us →

TOP_SIGNALS

OpenAI expands ChatGPT with MCP connectors for workflow automations, available now for Plus and Pro users

3,857 Likes

Anthropic introduces web fetch in Claude API for research, coding, and data extraction

3,328 Likes

Stability AI releases Stable Audio 2.5, generating three-minute tracks in seconds, built for enterprise audio workflows

748 Likes

ElevenLabs launches Voice Remixing to edit cloned and designed voices for agents

538 Likes

Anthropic confirms bugs caused degraded Claude responses, denies intentional degradation

3,538 Likes

TRENDING_PAPERS

Google introduces AI system to automate code generation for science

1,638 Likes

Google researchers address the bottleneck of slow scientific software development by introducing an AI system powered by Gemini. The system uses tree search and code rewriting to automatically generate, test, and refine algorithms.

OpenAI and Anthropic jointly study safety of Claude and GPT models

2,047 Likes

The joint study evaluated each other’s public models on safety, testing jailbreaks, hallucinations, and scheming, finding no model fully resistant but differences in robustness and refusal strategies.

Google study finds where embeddings fail

1,638 Likes

Google introduces the LIMIT dataset, showing single-vector embeddings collapse below 20% recall on dense two-document combinations, revealing fundamental limitations in handling concept mixing queries.

TOP_READ

Mira Murati’s lab finds cause of inconsistent LLM outputs

5,836 Likes

alpha_signal_image_4

ex-OpenAI CTO Mira Murati’s Thinking Machines Lab published its first research paper on reproducibility in LLM inference. The work identifies the real source of nondeterminism and proposes an open-source fix for consistent outputs.


Key takeaways from the research:

  • LLMs can produce different outputs even when temperature is set to 0.

  • Engineers often blamed floating-point math and concurrency for this issue.

  • The real cause is a lack of batch invariance, where results vary depending on how requests are batched on servers.

  • The team introduced batch-invariant kernels to make outputs consistent regardless of batch size or server load.

  • They released the solution as open source for community use.

  • Deterministic inference enables reproducible experiments, consistent business outputs, and true on-policy RL training.

Read now

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 250,000+ AI developers?

Work with us