<p><strong>TLDR:</strong> GPT-5.4 is leading CursorBench on correctness with efficient token usage, and OpenAI published the methodology behind how Cursor actually evaluates models. Meta’s AI position is looking increasingly shaky, with community sentiment turning sharply negative after reports they may license Gemini despite $20B in investment. Claude Code keeps shipping — voice mode on desktop, setup scripts, interactive charts in chat, and a new skill creator with A/B testing. Qodo is making noise with a code review benchmark showing a 12-point F1 advantage over Claude Code Review at 10x lower cost.</p>
<hr>
<h2>GPT-5.4 leads CursorBench</h2>
<p>OpenAI announced GPT-5.4 is at the top of CursorBench, Cursor’s internal benchmark for agentic coding tasks, leading on correctness while keeping token usage efficient. Cursor also published how CursorBench actually works: tasks are sourced from real Cursor sessions using Cursor Blame to trace committed code back to agent requests, prompts are intentionally underspecified, and tasks span multiple files — closer to real usage than SWE-bench. CursorBench-3 shows stronger model separation at the frontier where public benchmarks have saturated. Worth reading if you’re trying to make sense of which model to reach for.</p>
<h2>Meta’s AI standing collapses in community sentiment</h2>
<p>The news that Meta may license Gemini landed badly. The most-liked take today: “Meta’s AI efforts are wasted. They are even further behind than Google and xAI. Probably even behind Kimi K2.5. Only Anthropic and OpenAI will play the endgame.” A separate post with 32 likes eulogized Llama 1-3 as a genuinely good era for open source that’s now over. Whether Meta’s internal situation is actually that dire is hard to know, but the perception shift is real and fast.</p>
<h2>Claude Code ships voice mode, setup scripts, and a skill creator</h2>
<p>Anthropic pushed several Claude Code updates today. Voice mode is now live on desktop and Cowork. Setup scripts let you automate environment configuration before Claude Code launches. Claude can now generate interactive charts and diagrams directly in chat. The most interesting addition for power users is the skill creator plugin, which adds A/B testing for custom skills, parallel multi-agent test runs, and description optimization to improve trigger reliability — addressing the longstanding problem that skills often don’t fire when you expect them to.</p>
<h2>Qodo challenges Claude on code review</h2>
<p>Qodo published benchmark results claiming their extended mode beats Claude Code Review by 12 F1 points (74.7 vs 62.7), with identical precision at 79% but significantly higher recall (71% vs 52%), and at 10x lower cost. The methodology uses the same repos and judge for both tools. Take vendor benchmarks with appropriate skepticism, but the recall gap is large enough to be worth investigating if code review cost is a real concern for your team.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Gemini CLI plan mode:</strong> Google added a read-only plan mode to Gemini CLI that restricts the agent to read_file, grep_search, and glob before proposing changes — a sensible guardrail for large codebase exploration.</li>
<li><strong>VS Code Autopilot mode:</strong> VS Code’s coding agent now has an autopilot setting that runs tasks end-to-end without asking for confirmation at each step, alongside the existing default and bypass modes.</li>
<li><strong>Qwen 3.5 local performance:</strong> Community posts note Qwen 3.5 9B running locally is competitive with much larger paid models for coding tasks, with one developer switching from Claude/Gemini/OpenAI credits to Qwen Code and reporting strong results at zero cost.</li>
<li><strong>systemd adds <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a>:</strong> systemd 260-rc3 ships an <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> to guide AI coding agents on architecture and contribution style, plus a claude-review.yml workflow for PR review. AI-assisted contributions require disclosure.</li>
<li><strong>AMD Linux driver co-developed with Claude:</strong> AMD engineer Harry Wentland credited Claude Sonnet 4.5 as co-developer on color-space conversion patches for the AMDGPU driver, noting the AI “basically wrote all the code” while he steered and reviewed.</li>
<li><strong>Augment code review in production:</strong> A production comparison shows Augment-assisted review fixed 1.03 bugs per PR versus 0.54 for human reviewers, with comparable true-positive rates (45% vs 50%).</li>
<li><strong>Qt Creator 19:</strong> Released with a built-in MCP server letting AI tools like Claude Code open files, build, run, and debug directly inside the IDE.</li>
<li><strong>Context Mode MCP for Claude Code:</strong> A tool that indexes large tool outputs into a local SQLite FTS5 database instead of dumping them into the context window, reportedly extending usable session time from ~30 minutes to ~3 hours.</li>
<li><strong>MiniMax Agent Desktop:</strong> MiniMax shipped a free desktop agent tool positioned as a direct competitor to Claude Cowork, which sits behind Anthropic’s paid tier.</li>
<li><strong>OpenCode community PR quality:</strong> The OpenCode team posted a frank take that community PRs are “99% of the time slop” and consume significant team time to evaluate — a useful data point for any open-source maintainer thinking about AI-assisted contribution workflows.</li>
<li><strong>Codex Automations GA:</strong> OpenAI promoted Automations to general availability in the Codex app and shipped themes support including custom themes. Hooks are also coming to Codex.</li>
<li><strong>Kimi K2.5 at NVIDIA GTC:</strong> Moonshot AI founder Zhilin Yang is presenting “How We Scaled Kimi K2.5” at GTC on March 17, free to attend remotely.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

GPT-5.4 tops CursorBench, Cursor's internal agentic coding benchmark, with efficient token usage. OpenAI published CursorBench's methodology, which uses real Cursor sessions and multi-file tasks. Meta's AI credibility is taking hits as community sentiment turns negative following reports they may license Gemini despite $20B in AI investment. Claude Code shipped voice mode on desktop, setup scripts, interactive charts, and a skill creator with A/B testing. Qodo claims a 12 F1-point advantage over Claude Code Review at 10x lower cost. Other notable items include Gemini CLI plan mode, VS Code Autopilot, Qwen 3.5 local performance, systemd adding AGENTS.md, AMD crediting Claude as co-developer on a Linux driver, and Codex Automations reaching GA.

GPT-5.4 tops CursorBench, Meta's AI credibility takes another hit