<p><strong>TLDR:</strong> Claude Code shipped local scheduled tasks today, letting you set recurring background jobs directly from the desktop app. GPT-5.4 jumped six points on the Intelligence Index to tie Gemini 3.1 Pro Preview for the top spot, with a 1M token context window and strong agentic coding results. Qwen 3.5’s 4B model is reportedly matching GPT-4o quality while running on a phone. Meanwhile, a real-world reminder landed: Claude Code wiped a production database via Terraform, taking down 2.5 years of course submissions.</p>
<hr>
<h2>Claude Code ships scheduled tasks</h2>
<p>The <code>/loop</code> command and local scheduled tasks are now live in Claude Code desktop. You pass an interval and a prompt — <code>/loop 5m check if the deployment is healthy</code> — and Claude sets up a background job that fires on repeat for the duration of your session. This is a meaningful shift: agents running on a schedule while you’re doing other things is a different workflow than interactive coding. The community reaction has been positive, with several engineers noting that scheduling and automation are becoming a core part of how they work with coding agents.</p>
<h2>GPT-5.4 tops the Intelligence Index</h2>
<p>OpenAI’s GPT-5.4 jumped six points to tie Google’s Gemini 3.1 Pro Preview at the top of the Intelligence Index. The model ships with a 1M token context window and is drawing strong reviews for research-level tasks and agentic coding. Early hands-on reports are mostly positive — fast, verbose, good at multi-document analysis — though frontend UI work remains a weak spot and it still lags Claude and Gemini there. One notable regression: prompt injection via function call return data succeeds roughly 2% of the time. OpenAI also launched Codex for Open Source today, giving maintainers access to Codex capabilities, and Codex Security is now available on ChatGPT Pro accounts.</p>
<h2>Claude Code wiped a production database</h2>
<p>DataTalksClub published a full post-mortem: Claude Code executed a Terraform command that wiped their production database, taking down the course platform and destroying 2.5 years of homework submissions, projects, and leaderboards. Automated snapshots were gone too. This is the kind of story worth reading carefully if you’re letting agents touch infrastructure. The failure mode isn’t surprising — agents with Terraform access and insufficient guardrails — but the blast radius here was severe.</p>
<h2>Qwen 3.5 runs on phones, matches two-year-old frontier models</h2>
<p>Qwen 3.5’s 4B model is benchmarking at roughly GPT-4o quality, which came out in May 2024. It runs on modern mobile hardware. The implication being discussed: the gap between frontier datacenter intelligence and on-device intelligence might be running about two years behind, and closing. The 35B variant is available on Groq and getting traction as a cheaper alternative to Claude Pro for coding tasks, though at least one hands-on report from a WiFi-less flight found the local quant disappointing.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>T3 Code alpha:</strong> Theo and Julius shipped an open-source desktop app for managing parallel AI coding agents using worktrees, wrapping Codex CLI with Claude Code and Cursor support planned. Early stage, not accepting PRs yet.</li>
<li><strong>Cursor cloud agents:</strong> The Cursor team discussed their new cloud agents on the Latent Space podcast, including Composer 1.5 (RL-trained, between Sonnet 4.5 and Opus 4.5 in capability) and the infrastructure challenges of running millions of sandboxes for RL training.</li>
<li><strong>Cursor waiting on Anthropic:</strong> Cursor has a PR ready to add Claude support via the Agent SDK but is waiting for a thumbs-up from Anthropic leadership before shipping it.</li>
<li><strong><a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> research:</strong> ETH Zurich researchers tested four agents on 138 Python tasks and found LLM-generated <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> files decreased task success by ~3% and raised inference costs by over 20%. Human-written ones helped slightly but also raised costs ~19%.</li>
<li><strong>GitHub Copilot VS Code v1.110:</strong> Major update adds agent lifecycle hooks, conversation forking from checkpoints, auto-approval toggles, agent plugins, agentic browser control, and shared agent memory across Copilot tools.</li>
<li><strong>Figma MCP + VS Code:</strong> GitHub Copilot users can now connect to the Figma MCP server from VS Code for a bidirectional design-to-code workflow, with Copilot CLI support coming soon.</li>
<li><strong>Codex WebSocket drain:</strong> OpenAI is investigating reports of unexpectedly high usage drain in Codex when WebSockets are enabled.</li>
<li><strong>Prompt injection in GitHub Issues:</strong> A new attack surface identified where malicious content in GitHub Issue titles can manipulate AI agents doing automatic triage.</li>
<li><strong>LLM-generated code quality:</strong> A practitioner benchmarked an LLM-generated Rust SQLite reimplementation and found it 20,171x slower on primary key lookups due to a broken query planner and full fsyncs on every insert — a useful concrete example of plausible-looking but incorrect output.</li>
<li><strong>Lovable joins Anthropic Marketplace:</strong> Enterprise teams with existing Anthropic commitments can now deploy Lovable for non-engineering roles directly through the Claude Marketplace. Lovable also processes over 1 billion tokens per minute internally.</li>
<li><strong>GitHub Copilot execution:</strong> A post with 154 likes asking what happened to GitHub Copilot given their first-mover advantage is getting significant traction — no clear answer, but the sentiment is pointed.</li>
</ul>


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

Agentic Digest

A roundup of major AI coding tool developments: Claude Code now supports local scheduled tasks via a `/loop` command for recurring background jobs. GPT-5.4 tied Gemini 3.1 Pro Preview at the top of the Intelligence Index with a 1M token context window and strong agentic coding performance. A notable incident saw Claude Code wipe a production database via Terraform, destroying 2.5 years of course data. Qwen 3.5's 4B model reportedly matches GPT-4o quality and runs on mobile hardware. Additional items cover Cursor cloud agents, GitHub Copilot VS Code v1.110 updates, AGENTS.md research showing LLM-generated files hurt task success, prompt injection risks in GitHub Issues, and LLM-generated Rust code running 20,000x slower than expected.

Claude Code gets scheduled tasks, GPT-5.4 lands at the top of the Intelligence Index

<p>The loop command is so exciting. Gonna try that out</p>
