DAILY DIGEST
Google says Antigravity’s Voice Mode lets developers steer its coding agent and approve some permissions by voice; it is rolling out gradually in App 2.29 and CLI 1.5.0+. Remote Agents run in isolated cloud VMs and are available to Google AI Pro and Ultra subscribers, with phone control; the update also redesigns the interface and Automations page.
If you use Antigravity, check for the rollout and try remote runs from your phone.
Anthropic describes agent memory as readable notes retained between sessions, not model-weight changes. Its offline “dreaming” process reviews transcripts and notes to correct errors, remove stale items and reorganize memory. Anthropic reports Opus 5.5 improved from 0.644 to 0.833 on CL-Bench with memory; on a wiki task, scores across 25 questions rose from 13% to 48% when dreaming scaled from 2 to 128 sessions.
Consider memory and offline consolidation for agents that need continuity across long-running tasks.
Alibaba researchers’ Co-Design Bench covers 273 multimodal tasks drawn from real coding sessions. The strongest tested endpoint, GPT-6 Astra, fully completed 26.4% of tasks and met 69.9% of acceptance criteria on average; full completion fell to 17.9% for repository-wide edits. These are benchmark results, not a guarantee of performance on your codebase.
Evaluate agents on complete task success, especially for changes spanning a repository.
Google DeepMind says supervised AI agents found two weaknesses in legacy 1024-bit RSA. It estimates a signature-forgery attack is about 1,600 times faster than the prior best, though it requires billions of signing queries and does not recover the private key. A separate theoretical batch-factoring approach could lower the cost per additional key after expensive precomputation; 2048-bit-and-larger keys are unaffected, according to the report.
Check for any remaining 1024-bit RSA dependencies; the report says modern key sizes are unaffected.
OpenAI announced a partnership with security firm Hacktron to search for vulnerabilities in its systems, focusing on interactions with third-party platforms. OpenAI engineers and Hacktron researchers will choose priority areas together; Hacktron receives expanded access to frontier models through OpenAI’s Daybreak Red program. The captured report gives no findings or timeline.
Treat this as a security-testing initiative, not a disclosed vulnerability or remediation notice.
Bloomberg reported that DeepSeek said its open-source agent app Harness reached 1.54 million daily active users on November 4, less than three months after its August launch. The reported usage was about 92 trillion tokens a day; the count covers users connecting through DeepSeek’s official API. The capture gives no pricing or coding-specific performance data.
Treat the user count as adoption context, not evidence of agent quality or API economics.
Axiom Math announced that AI agents translated the roughly 7,654-page Stacks Project into about 16 million lines of computer-checked code, formally verifying the mathematics. The company says it aims to build a verified foundation for AI agents in software and science; the captured report provides no independent validation details or developer release information.
Watch for tools or methods that make formal verification practical in software projects.
Meta began rolling out specialized Muse and Meta AI models trained using its Muse Wonder model. The company says Active Memory improves recall across conversations and Active Focus helps models persist through task obstacles; these features are rolling out to all users. The larger Muse Wonder Full model remains available to paid subscribers, with no developer API details in the capture.
This is a consumer-facing model update; don’t assume the features are available through an API.
Updated 2026-11-09 13:05 UTC
A compact digest when Nick publishes. No extra noise.