Today in AI: agents get graded like teammates, and game worlds start talking back

Oct 27, 2026

NEW: Cursor's design mode gets Annotate, so you can point at your UI instead of describing it

Cursor added Annotate to its design mode on Oct 27. You can edit elements, draw on the page and review your design at different screen sizes, and those marks go to the agent as visual prompts. Press Cmd+Shift+D in the browser to start. For AI coders doing front-end work, a circle and a short note replaces a paragraph like "move the second button 8px left on mobile", which is how designers already give feedback to people.

Source: https://x.com/cursor_ai/status/2115159585548021814

Oct 27-28, 2026

BENCHMARKS: FrontierCode 2.0 asks if you should merge the agent's PR, and it now anchors a new model leaderboard

Cognition launched FrontierCode 2.0 on Oct 27. Its tasks come from real open-source repos, each built by a maintainer of that repo with more than 40 hours of work, and the prompts are deliberately underspecified so models have to make engineering calls. Grading asks whether a maintainer would, and should, merge the change. Cognition says the top model scores just 45%, and GPT-6.1 Sol leads on cost efficiency at 39% for $1.03 per task. Hours later, Artificial Analysis folded it into Intelligence Index v5, where 60% of the score now comes from private test sets. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max) at 57, GPT-6.1 Sol (max) at 55 and Gemini 4 Argon (high) at 50. Sol scores within 4 points of Opus at about an eighth of the cost per task ($1.27 vs $10.04). The gap to open weights widened from 3 points to 12: Kimi K3.1 Pro scores 46 but its weights are not out yet, and the best released-weights models, DeepSeek V4.1 Flash and GLM-5.3, score 32. If you pay per token, Sol is worth a test run, and local models are further behind than older leaderboards suggested.

Source: https://x.com/ArtificialAnlys/status/2115265133467635877

Oct 27, 2026

WORLD MODELS: Google DeepMind's Genie 4 generates explorable worlds with sound and talking characters

Google DeepMind announced Genie 4 on Oct 27, a real-time interactive world model built on Gemini Omni, with Gemini 3.8 Flash acting as a guiding agent that keeps the scene coherent. You move with standard keyboard controls and steer with text, like changing the weather or adding a character. It is DeepMind's first real-time world model to generate audio natively, including ambient sound and dialogue. The Verge reports sessions of about two minutes, up from one for Genie 3. It is a research preview for trusted testers for now, with Google AI Ultra subscribers (in Project Genie) and early Gemini API and AI Studio partners next, and a waitlist is open. DeepMind lists the limits itself: it stays coherent only over short sessions, details can drift over time, and legible text is still hard. For indie game devs it is not an engine, but it could become a fast way to test the feel of a scene or mechanic before you build it.

Source: https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/

Oct 27-28, 2026

GAME DEV: Epic says it will lift the Unreal Engine ban on AI training on engine output

Epic CEO Tim Sweeney said that Unreal Engine's EULA currently prohibits training AI on the output of Unreal Engine and MetaHuman, and that an upcoming EULA change removes that restriction for Epic's own engine output. Each developer still releases their game under a license of their choosing, which can allow or forbid AI training, and content on Fab follows each creator's NoAI tag. Fortnite island creator content is not affected. Sweeney says the benefit is that AI models will get more familiar with Unreal and better at helping with code and content in UE projects. Separately, Epic's Marcus Wassmer confirmed that Epic uses AI in Unreal Engine development. This follows Kotaku's report on Microsoft and Epic takedowns of AI-made game recreations. The new EULA text is not published yet, so check the final terms before you rely on it.

Source: https://x.com/TimSweeneyEpic/status/2115234757554672065

Oct 27, 2026

AGENTS: Experts went with their agent's hidden lean 86% of the time

In a randomized study published Oct 27, Anthropic gave 50 professional analysts a data analysis task in Claude Code with four judgment calls that each had two reasonable answers. A coin flip secretly set which way each analyst's agent leaned, and the agent was told to argue honestly for it. Analysts went with the agent's lean 86% of the time, against about 50% if it had no effect. When they had already stated a preference and the agent disagreed, they switched 60% of the time. Afterward, 77% explained their choices with the agent's own arguments, but only 15% credited the agent. Analysts using the agent were also faster and made far fewer serious errors. Anthropic notes it strengthened the leans on purpose and tested one task with one model. For AI coders, the practical fix is simple: when an agent recommends a library, schema or architecture, decide what you think first, or ask it to make the best case for the other option too.

Source: https://www.anthropic.com/research/agents-judgment-calls

Oct 26, 2026

PLATFORM: Former GitHub CEO ships Entire, a Git forge built for agents

Thomas Dohmke launched Entire on Oct 26 as "the Git forge for agent-human collaboration". Trails replace the usual pull request loop: evals feed the next round of agent work, and sessions, code changes and CI checks stay together. Continuous Memory stores agent session logs in the repo as Git objects, and Entire Graph and Entire Brain share context across agents with what Entire says is up to 71% fewer tokens. Hosting was rebuilt for the volume of changes agents produce, and you can mirror from GitHub and move when ready. It works with Claude Code, Codex, Cursor, GitHub Copilot, Kiro and others. The token figure is Entire's own. Entire raised a $60M seed at a $300M valuation in February.

Source: https://x.com/ashtom/status/2114763996691562540

Oct 27, 2026

DEALS: Infisical reaches a $600M valuation keeping secrets away from agents, and Hex raises $100M to go agent-native

Infisical raised a $40 million Series B at a $600 million valuation on Oct 27, led by Elad Gil, with Y Combinator and Gradient Ventures also in (founder post). Its Agent Vault lets companies give each agent only the access it needs without handing it the credentials, so keys cannot leak in an attack or get sent to an LLM provider. Infisical says it serves more than 1,300 customers and 300,000 developers, including Hugging Face and OpenRouter, and secures 10 billion secrets a day. Also on Oct 27, Hex closed $100 million at a $1.2 billion valuation, led by Peak XV and 137 Ventures with a16z, Sequoia and Redpoint participating, and launched Hex 4.0: a from-scratch rewrite of its data platform for the era of agents, built with agents writing the code and designed to work natively with outside agents instead of only through MCP. Both rounds back the same idea: as agents do more of the work, the tools that keep them safe and grounded are worth a lot.

Source: https://www.prnewswire.com/news-releases/infisical-raises-40-million-at-600-million-valuation-to-secure-the-ai-agent-economy-302920994.html

Ship faster this week

Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.