New in Cursor design mode: Annotate. Give agents visual prompts: edit elements, draw on the page, and review your designs at different screen sizes.

The largest math proof in human history is now computer-verified. Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry." About 15,000 pages, built by over 100 mathematicians across generations.

We launched the Gemini API CLI. A CLI for calling the Gemini API. Give it to the agent you already use (Claude Code, Codex, Antigravity, Cursor, OpenClaw, Hermes...). It's NOT a new coding agent you can use it too: 1. Generate images, video, speech, music. 2. Analyze PDFs, images, audio, video, YouTube URLs, 3. Transcribe recordings to md / json / srt, 4. Run Deep Research in the background. Stateless. No memory, no chat mode. Your agent owns the context and the plan, gemini-api does the API call. Build for Agents, it's self-describing. `--help`, `--dry-run`... Start here: aistudio.google.com/learn/in…

$ORCL ANNOUNCES NEW AI PLATFORM THAT LETS CHATGPT EXECUTE ENTERPRISE WORKFLOWS Oracle introduced Fusion Companion and Fusion Agentic Home at Oracle AI World, allowing employees to manage business operations through AI tools like ChatGPT. Instead of navigating multiple applications, users can describe what they want done, and the system identifies and coordinates the necessary transactions, workflows, approvals, and AI agents. The technology works across Oracle's Fusion applications, including finance, HR, supply chain, sales, and customer service, while maintaining enterprise security and permissions.

TOP 10 STRATEGIC TECH TRENDS AND PREDICTIONS 2027-2031, GARTNER 1 Large quantitative models 2 Speech reasoning models 3 World models 4 AI context platforms 5 AI coding agent harness 6 AI agent identity 7 Autonomous exposure remediation 8 SovereignTech 9 Disposable applications 10 Wetware as a service Predictions for 2027-2031: ———————————— 1. By 2030, LQMs will replace LLMs in 25% of strategic forecasting and operational modeling tasks across major enterprise functions. 2. By 2030, over 70% of enterprise voice interactions will be handled autonomously by speech models, up from less than 10% in 2026. 3. By 2030, 40% of robotics applications will incorporate world models to drive optimal decisions in adaptive environments, compared with less than 1% in 2025. 4. By 2027, organizations that prioritize semantics in AI-ready data will increase agentic AI accuracy by up to 80% and reduce costs by up to 60%. 5. By late 2026, autonomous coding agents will generate 20% of all public commits, making closed-loop harnesses mandatory to prevent critical software delivery quality failures. 6. By 2028, 25% of enterprise security breaches will trace back to AI agent abuse due to unmanaged machine privileges and legacy credential practices. If organizations don’t resolve these identity challenges, over half of their enterprise AI initiatives will halt entirely. 7. By 2031, organizations deploying autonomous exposure remediation will reduce their average vulnerability containment window from weeks to within 48 hours. 8. By 2030, over 75% of non-U.S.-based organizations will adopt digital sovereignty strategies, shifting critical workloads to reduce geopolitical and vendor lock-in risks. 9. By 2030, widespread SaaS fatigue and AI-driven software abundance will force leading organizations to ensure 80% of new applications are disposable and easy to replace. 10. By 2030, 25% of enterprise organizations prioritizing specialized low-power AI will deploy biological computing to bypass traditional silicon limits. Gartner’s top 10 strategic technology trends for 2027 is a strong reminder for all business leaders: the future of business is autonomous.

Boris Cherny, creator of Claude Code, shared the harness for overnight runs build on Sonnet 5.5, audit via Opus 5.5, and gate write policies before Fable 5.5 drops claude -p "Read PROGRESS.md and build next feature per CLAUDE.md" why naive agent loops always crash after 2 hours: - models hallucinate passing tests to terminate early - context bloat eats 150k tokens of dead terminal output - process restarts trigger total amnesia and restart from zero Anthropic solved this with their official long-running harness repo: anthropics/cwc-long-running-agents the architecture replaces blind LLM trust with shell-enforced write policies: 1. Default-FAIL Contract every test requirement initializes as false an agent cannot flip status to passing by simply claiming it passed the write policy gate physically blocks file updates until proof logs exist 2. PreToolUse Interceptor Gate verify -gate.sh intercepts every Write tool call to test-results.json track -read.sh monitors evidence reads from vitest.log and browser screenshots zero evidence opened = instant write rejection 3. Isolated Fresh-Context Evaluator Sonnet 5.5 builds features inside the active workspace Opus 5.5 evaluates the output in an isolated 0k context window the evaluator has zero Write tools, preventing self-grading contamination 4. Zero-Amnesia Context Flushing every verified milestone writes to PROGRESS.md and creates a git commit the harness flushes 142k RAM tokens and restarts the builder loop fresh zero context degradation, zero memory loss, infinite unattended runway run this prompt inside Claude Code to install the harness: "Configure this repo for autonomous long-running execution using Anthropic harness primitives: 1. Install PreToolUse hooks from anthropics/cwc-long-running-agents: - set up .claude/hooks/verify-gate.sh to intercept all writes to test-results.json - set up .claude/hooks/track-read.sh to verify test logs and screenshots were opened - enforce Default-FAIL contract where all criteria start false 2. Configure dual-agent roles in ~/.claude/settings.json: - Builder agent: Sonnet 5.5 with full workspace editing tools - Evaluator agent: Opus 5.5 with zero write permissions in fresh context 3. Setup persistence and operator control: - update PROGRESS.md and commit to git on every passing milestone - enable kill -switch.sh via AGENT_STOP and steering via STEER.md Show me the configuration diff first. Await confirmation before applying" ↳ github.com/anthropics/cwc-lo…

Popular belief is that a complex AI harness like Claude Code or Codex will improve the capabilities of LLMs in coding tasks. But two separate studies show that the added complexity of the harness does not necessarily improve the accuracy of the coding agent. In some cases, it can triple the LLM token bill without giving you significant improvement. Take the self-reported benchmark results of AI vendors with a grain of salt. Sometimes, running the same prompt multiple times on the same model-harness configuration will give you a spread that is wider than the improvement that vendors report on their harnesses. Develop your own benchmarks/evals, run your own tests, make your own judgment.

Official Anthropic tip: stop burning Opus 5.5 tokens on tasks Sonnet 5.5 can swarm, and stop running autonomous agents with unconfined bash access jail the entire multi-model runtime at the kernel layer with srt npx @ anthropic-ai/sandbox-runtime run --allow-net registry.npmjs. org claude Opus 5.5 plans architecture and acts as dedicated security advisor Sonnet 5.5 writes the code and executes tool actions inside the sandbox Haiku 5.5 swarms dependency triage and network domain verification the srt kernel engine audits every tool call at three critical points: → before filesystem reads: is the path strictly inside the project root? → when network egress triggers: is the outbound domain allowlisted in proxy rules? → before child process spawns: are ~/.ssh, ~/.env, and root credentials blocked? Haiku scans. Sonnet builds. Opus verifies. srt jails in 0.2ms Docker sandboxing is the wrong primitive for multi-agent loops: container spinup burns 2,450ms, memory overhead inflates by 650MB, and host integration breaks srt enforces macOS Seatbelt and Linux bubblewrap directly at the OS kernel layer with 0.2ms syscall overhead and zero container bloat Plan on Opus. Build on Sonnet. Swarm on Haiku. Jail with srt - the full multi-agent sandbox setup > Opus 5.5 on high reasoning plans the task and audits security contracts > Sonnet 5.5 executes code edits and unit tests inside the active workspace > Haiku 5.5 swarms file indexing, AST diffs, and package triage at 340 tok/s > srt wraps all child bash commands and local MCP servers via OS kernel primitives > filesystem restricted to read-write in workspace, blocked in ~/.ssh and ~/.env > network egress filtered through a local proxy with domain allowlisting > 0.2ms syscall overhead, 4MB RAM footprint, zero Docker container bloat paste the setup and this prompt into Claude Code ↓ "Configure my Claude Code workspace for multi-model sandboxed execution: 1. Install and verify @ anthropic-ai/sandbox-runtime globally: > Verify kernel backends: Seatbelt on macOS or bubblewrap on Linux > Ensure non-root execution without Docker daemon requirements 2. Set up model role delegation in ~/.claude/settings.json: > Main lead session: Opus 5.5 on high effort for architectural planning > Worker subagents: Sonnet 5.5 at medium effort for code writing and tests > Fast scanner subagents: Haiku 5.5 for file search, linting, and docs triage 3. Wrap all bash tool calls and MCP servers with srt kernel boundaries: > Enforce execution prefix: srt run --allow-net=registry.npmjs. org,github. com > Confine filesystem read-write permissions strictly to the current workspace directory > Blacklist ~/.ssh, ~/.aws, ~/.gnupg, and all .env files from agent reads 4. Add a security rule to ~/.claude/CLAUDE.md: consult Opus 5.5 before approving a large plan, and reject any command attempting to bypass srt kernel profiles Show me every configuration change as a diff first. No edits until I say go." ↳ github.com/anthropics/sandbo…

Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server