Key Signals

• OpenAI Codex 0.120.0 ships Realtime V2 background agent streaming, marking a significant step toward real-time collaborative agentic workflows. The release enables background agents to stream progress while running and queue follow-up responses, moving Codex closer to a truly interactive multi-agent coding experience. Code-mode tool declarations now include MCP outputSchema details for typed structured tool results, tightening the integration between Codex and the Model Context Protocol ecosystem. This release also addresses critical sandbox and permission bugs on Windows, improving cross-platform reliability for enterprise environments. [1]

• The Linux kernel has adopted a formal AI Coding Assistants Policy, establishing the Assisted-by attribution tag as a standardized format for tracking AI contributions to the world's most important open-source project. The policy explicitly prohibits AI agents from adding Signed-off-by tags — only humans can certify the Developer Certificate of Origin — and requires human submitters to review all AI-generated code and take full legal responsibility. The tag format (Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2]) sets a precedent that other large open-source projects are likely to adopt, creating a traceable record of AI's evolving role in kernel development. [4]

• Andrej Karpathy's "AI Psychosis" essay captures a widening perception gap: professional developers using frontier AI coding tools like Codex and Claude Code are experiencing transformative productivity gains, while casual users remain skeptical. The gap exists because software is the first domain where frontier-model capability, AI fluency, and deep domain expertise all converge in the same people. Meta's internal "Claudeonomics" leaderboard — where 85,000+ employees consumed 60 trillion tokens in 30 days — illustrates how aggressively large engineering organizations are leaning into AI-assisted development. As Anthropic moves Claude Cowork to GA for enterprise users beyond developers, this dynamic is poised to spread into law, finance, and operations. [3]

• Goldman Sachs data reveals an AI productivity paradox: the five largest US tech companies will spend $667 billion on AI infrastructure in 2026, yet measurable productivity gains remain narrowly concentrated in software development and customer support (~30% median improvement). UC Berkeley researchers found that AI doesn't reduce workloads but intensifies them through "workload creep" — tasks get faster, expectations rise, and scope expands until cognitive fatigue degrades decision quality. 62% of associates and 61% of entry-level workers report AI-related burnout, compared to only 38% of C-suite executives, revealing a structural gap between AI strategy-makers and the people managing its daily outputs. [6]

• ProPublica's journalists staged the first major US newsroom strike partly demanding protections from AI-related layoffs, filing an unfair-labor-practice charge with the NLRB over unilateral AI policy implementation. The 24-hour walkout by roughly 150 ProPublica Guild members signals that AI labor conflicts are expanding beyond tech into knowledge-work industries. The New York Times Guild is separately negotiating AI contract language including revenue-sharing when journalist work is licensed for AI training, establishing bargaining precedents that could reshape how organizations deploy AI across creative and analytical roles. [5]

AI Coding News

• Andrej Karpathy's essay on "AI Psychosis" identifies software development as the first domain experiencing transformative AI impact, and the perception gap is widening fast. Karpathy argues that reinforcement learning works best with verifiable reward functions, making coding and math dramatically more trainable than other domains — and because these generate the most B2B revenue, that's where the biggest teams are focused. The article also covers Anthropic's Claude Mythos Preview, which developed 181 working exploits for Firefox JavaScript engine vulnerabilities compared to just 2 for the prior Opus 4.6 model — capabilities that "emerged as a downstream consequence of general improvements in code, reasoning, and autonomy." Goldman Sachs estimates AI is eliminating roughly 16,000 net US jobs per month while a Gallup study found 31% of Gen Z now says AI makes them angry, up 9 points from last year. [3]

• The Linux kernel has published an official AI Coding Assistants Policy requiring an Assisted-by attribution tag for all AI-assisted contributions. AI agents are explicitly prohibited from adding Signed-off-by tags, as only humans can legally certify the Developer Certificate of Origin. The standardized format — Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2] — tracks the specific AI tool, model version, and optional analysis tools like coccinelle or sparse. All AI-generated code must comply with GPL-2.0-only licensing, and human submitters bear full legal responsibility for reviewing and certifying every contribution. [4]

• ProPublica's journalists staged the first major US newsroom strike partly over AI-related protections, with roughly 150 members of the ProPublica Guild walking out for 24 hours. The union is seeking contract language that would prohibit AI-driven layoffs and filed an unfair-labor-practice charge with the NLRB after ProPublica published AI editorial guidelines without bargaining with staff. The New York Times Guild is also negotiating AI provisions including a proposal for Guild members to share in revenue when their work is licensed for AI training. This precedent could ripple across knowledge-work industries where AI is reshaping workflows. [5]

• AI adoption is intensifying work rather than reducing it, with UC Berkeley researchers identifying "workload creep" as a core mechanism and BCG coining the term "AI brain fry." An eight-month study at a 200-person tech firm found that as tasks got faster, expectations and scope expanded — product managers started writing code, researchers took on engineering work. BCG found that 14% of AI power users experience mental fog, difficulty focusing, and headaches after extended tool interaction. Goldman Sachs reported no meaningful economy-wide relationship between productivity and AI adoption, despite record S&P 500 AI discussion on earnings calls, with measurable gains confined to software development and customer support at roughly 30%. [6]

Feature Update

• OpenAI Codex 0.120.0 introduces Realtime V2 background agent streaming and MCP outputSchema support for typed structured tool results. Background agents can now stream progress while still running and queue follow-up responses until the active response completes, enabling more fluid multi-turn agentic workflows. Hook activity in the TUI has been redesigned with live running hooks shown separately and completed output retained only when useful. SessionStart hooks can now distinguish sessions created by /clear from fresh startup or resume sessions. Critical bug fixes address Windows elevated sandbox handling for split filesystem policies, symlinked writable root permissions, codex --remote wss:// panics via Rustls crypto provider installation, and MCP cleanup on disconnect. Guardian follow-up efficiency improves by sending transcript deltas instead of full history. [1]

• Gemini CLI v0.39.0-nightly.20260411 delivers stability improvements including memory leak resolution, OOM prevention, and PTY exhaustion fixes. Lifecycle memory leaks from listeners and root closures have been cleaned up, and a buffer slice removal prevents out-of-memory crashes on large output streams. Orphan MCP subprocess leaks are now resolved, addressing a significant reliability issue for long-running sessions. New features include persisting subagent agentId in tool call records for better traceability, Strategic Re-evaluation guidance added to the system prompt, and increased codebase investigator turn limits to 50. Windows sandbox initialization performance is improved via native ACL application, and the build switches credential storage from keytar to @github/keytar. [2]