EN RU

Daily digest

9 items · ~9 min · Week 2026-W33

Must-read (2)

Ouroboros: A Self-Developing Coding Agent That Evolves Its Own Harness Through Reviewed Commits

Research official 2 src. ~1 min

Ouroboros is a coding-agent harness whose tools, prompts, context assembly and core implementation evolve through reviewed commits, combining a recursive free-evolution loop with experience-driven changes surfaced by ordinary usage. Running on Opus 5, it reports 86.74% on Terminal-Bench 2.1 and 90.69% on OSWorld-Verified, both new best-reported results, plus SOTA on a five-rollout CL-Bench campaign.

Why it matters
Received 1.1k upvotes on HuggingFace Daily Papers, reflecting strong community interest; a concrete data point in the self-improving-agent direction, including an explicit discussion of the safety problem of agents that rewrite their own code.
#agentic #self-improvement #coding-agents

BDH-CQ: Recurrent Latent Reasoning Model Breaks ARC-AGI-1 Cost-Efficiency Frontier

Pathway
Research official 2 src. ~1 min

BDH-CQ combines in-context learning with recurrent latent reasoning: it updates a recurrent memory from inputs given at inference time and solves queries through iterative computation in latent space, without producing a verbalized chain-of-thought. A 150M-parameter configuration reaches 29.5% pass@2 on the public ARC-AGI-1 evaluation set at $0.0007 per task, beating the previously reported cost-accuracy Pareto frontier for that benchmark.

Why it matters
Received 681 upvotes on HuggingFace Daily Papers; shows a non-chain-of-thought, small-model path to competitive ARC-AGI performance at very low inference cost, relevant to the ongoing debate on whether verbalized reasoning is necessary for strong performance.
#reasoning #latent-reasoning #efficiency #arc-agi

Worth knowing (3)

Claude Research Model Raises Riemann Zeta Zero Lower Bound from 41.6% to 67.2%

Anthropic
Research official + media 2 src. ~1 min

An unreleased Anthropic research version of Claude, working across two Claude Code sessions with about 60 sub-agents and 31M output tokens, improved a decades-old lower bound on the fraction of Riemann zeta zeros satisfying the Riemann hypothesis, combining recent number-theory results with Bombieri's 2000 work into a new Weil-quadratic-form argument. Two Anthropic mathematicians and outside experts (Brian Conrey, Dan Goldston) reviewed the result, and Claude also produced a Lean-verifiable version of the proof.

Why it matters
It does not prove the Riemann hypothesis and Anthropic does not expect the technique to lead there, but it is a rare case of a frontier model producing an expert-reviewed advance on an open problem in pure mathematics, illustrating both the promise and current limits of agentic AI-assisted research.
#reasoning #mathematics #agentic-research

OpenAI Expands Daybreak Cybersecurity Program, Launches GPT-5.6-Cyber

OpenAI
Tools official + media 2 src. ~1 min

OpenAI expanded its Daybreak initiative for vetted security teams, adding a specialized GPT-5.6-Cyber model (built on GPT-5.6 Sol) for authorized vulnerability research and exploit validation, alongside two access tiers: Daybreak Blue for general-purpose defensive work and Daybreak Red for purpose-trained cyber models.

Why it matters
Signals a shift toward gated, capability-tiered access to frontier models for dual-use security tasks, as labs balance defender empowerment against misuse risk.
#cybersecurity #api #gpt-5.6

xAI Launches Grok Bot Public Beta, Always-On AI Agent Teammates

xAI
Tools official + media 2 src. ~1 min

xAI opened a public beta of Grok Bot, AI agents that get their own cloud computer, sign into a user's existing apps and tools, retain context across tasks, and carry out multi-step work autonomously, checking back only when approval is needed.

Why it matters
Pushes xAI further into the agentic-AI/autonomous-coworker space where it now competes directly with Anthropic and OpenAI's agent products.
#agents #grok #automation
For reference (4)

VK Unifies User Data Across Services into a Single AI Neuro-Profile for Recommendations

VK
Industry official + media 3 src. ~1 min

VK combined anonymized user actions across VK Video, VK Clips, and VKontakte into a single transformer-based 'neuro-profile' to improve content recommendations, reporting a 5.5% increase in VK Clips watch time and higher engagement in testing.

Why it matters
Shows a major Russian platform consolidating cross-service behavioral data into one AI model to boost engagement and ad targeting.
#recommendation-systems #transformer #russia

Anthropic Extends Compliance API to Claude Cowork and Claude Code

Anthropic
Tools official 1 src. ~1 min

Anthropic expanded its Compliance API (beta, for Claude Enterprise customers) to cover Claude Cowork and Claude Code across desktop, web, mobile, and CLI, returning consolidated session transcripts with prompts, responses, and tool activity for audits and eDiscovery.

Why it matters
Gives enterprise compliance teams unified programmatic visibility across Anthropic's agentic products, using existing Compliance Access Keys without separate integration work.
#compliance #enterprise #claude-code #claude-cowork #api

GitHub Copilot CLI v1.0.79 Adds Kimi-K3 Support and Reworks Sandbox Auth Settings

GitHub (Microsoft)
Tools official 1 src. ~1 min

GitHub Copilot CLI v1.0.79 adds support for the kimi-k3 model, switches large monorepos to trigram-indexed tgrep instead of ripgrep, reworks sandbox settings by moving git/gh auth under a new sandbox.auth.* namespace with no migration path, and renames allowDevToolCaches to allowDevToolAccess as a breaking change.

Why it matters
The breaking sandbox-setting renames mean existing managed/MDM policies silently revert to defaults unless updated, which can quietly widen tool access for teams that don't catch the changelog.
#github-copilot #cli #coding-agent #sandbox

Claude Code v2.1.228 Hardens Skill Sync and Fixes Remote Control Session Leak

Anthropic
Tools official 2 src. ~1 min

Claude Code v2.1.228 hardens skills synced from claude.ai so they can no longer shadow local commands or MCP prompts, fixes a Remote Control /resume bug that leaked a resumed conversation's title/history into a connected session, and improves Vertex AI credential handling to fail fast on expired credentials instead of retrying for minutes.

Why it matters
Security hardening of synced skills and the Remote Control fix close real information-leak and impersonation risks for teams running Claude Code across multiple machines.
#claude-code #cli #coding-agent