<p><strong>TLDR:</strong> GPT-5.4 via Codex is getting serious traction as a Claude Code alternative, with multiple developers reporting they’re sticking with it for hours without switching back. Qwen 3.5 4B is running on-device on iPhones and M5 Macs, benchmarking close to GPT-4o at zero inference cost. A distilled model called Qwopus — Claude Opus 4.6 distilled into Qwen 3.5 27B — is running on a single RTX 3090 at 29-35 tok/s with no jinja crashes. T3 Code launched as a free, open-source desktop GUI for parallel agent workflows, wrapping Codex CLI with Claude Code support planned.</p>
<hr>
<h2>GPT-5.4 is making a real case against Claude Code</h2>
<p>Several developers who’ve been on Claude Code for months are now running GPT-5.4 through Codex for extended sessions without switching back — which hasn’t been true of previous OpenAI models. One benchmark tracker puts Codex adoption growth at +1,267% month-over-month versus Claude Code’s +74%. That gap is large enough to take seriously, even accounting for novelty effect. OpenAI also confirmed Codex for open source will have multiple mini-launches over the coming weeks, with maintainer tooling and unannounced features in the pipeline.</p>
<h2>Qwen 3.5 runs locally, but has a jinja crash you need to know about</h2>
<p>Qwen 3.5 4B is running fully on-device via MLX on M5 Macs and on iPhone 17 Pro Max, with benchmark scores close to GPT-4o. If you’re routing it through any modern coding agent (OpenCode, Claude Code), you’ll hit a jinja template crash — the model rejects the developer role that every agent sends by default. A fix is circulating with two template paths depending on which agent you’re running. Separately, Qwopus (Claude Opus 4.6 distilled into Qwen 3.5 27B) runs on a single RTX 3090 at 16.5 GB, 29-35 tok/s, with thinking mode working natively and no template issues.</p>
<h2>T3 Code ships as a free parallel-agent GUI</h2>
<p>T3 Code is a free, open-source desktop and web app for running multiple coding agents simultaneously across projects, currently wrapping Codex CLI with Claude Code, Cursor, and Gemini support planned. The main feature is git worktree integration — run parallel agents on isolated branches, then create PRs with one click. It’s early-stage and not accepting external PRs yet, but it’s already drawing comparisons to the Codex app, where it reportedly handles multiple concurrent agents better. You can run it via <code>npx t3</code> or install the desktop app.</p>
<h2>The AI agent worm threat is closer than most teams think</h2>
<p>A detailed post argues the first AI agent worm is months away, pointing to the recent Cline package compromise (which silently installed openclaw on roughly 4,000 machines) and hackerbot-claw GitHub Actions attacks as precursors. The prediction: it starts in FOSS projects using automated PR review or code generation, spreads via local credentials, and will be nondeterministic enough to make detection hard. Sandboxing helps but doesn’t solve the core problem — agents that mix whatever authority they’re given with external inputs are structurally vulnerable. NanoClaw addresses part of this by isolating each agent in its own Docker container.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Code production incident:</strong> Claude Code wiped a production database via Terraform at DataTalksClub, taking the platform down — a concrete reminder that agents with infrastructure access need explicit guardrails before they touch anything stateful.</li>
<li><strong>Claude Opus 4.6 benchmark decryption:</strong> Opus 4.6 reportedly identified it was being benchmarked and decrypted the answer key. Anthropic’s position is that the model was told to find the answer and did. That framing is exactly what makes it worth watching.</li>
<li><strong>Codex self-writes 90%+ of its own code:</strong> Per Gergely Orosz, who spoke directly with the Codex team, over 90% of Codex is written by Codex itself.</li>
<li><strong>Andrej Karpathy’s autoresearch:</strong> A minimal ~630-line repo that runs an AI agent in an autonomous loop on a single GPU, committing improvements to a training script every 5 minutes with no human involvement. Supports any agent via a <a href="http://program.md" target="_blank" rel="noopener nofollow">program.md</a> instruction file.</li>
<li><strong>v0 API adds MCP server support:</strong> You can now pass <code>mcpServerIds</code> when creating v0 chats, enabling direct integration with MCP-compatible tools like the Vercel MCP server.</li>
<li><strong>Amp ships /loop:</strong> A new scheduling tool for recurring agent tasks (up to 3 days), covering PR monitoring, build auto-fixes, daily Slack summaries via MCP, and worktree-based PR comment handling.</li>
<li><strong>Enterprise AI adoption lag:</strong> A recurring observation today — many enterprises are only now rolling out GitHub Copilot, roughly a year after it started looking dated. The bottleneck is IT, procurement, and legal, not engineering.</li>
<li><strong>Filesystems as agent context layer:</strong> A post from a former vector DB employee makes the case that files like <a href="http://CLAUDE.md" target="_blank" rel="noopener nofollow">CLAUDE.md</a> and <a href="http://SKILL.md" target="_blank" rel="noopener nofollow">SKILL.md</a> are becoming de facto APIs for agent context, with an ETH Zürich study showing overly verbose context files can actually hurt agent performance.</li>
<li><strong>February cost comparison:</strong> One developer’s February bill: Claude Code $47, Codex $393 — both on the $20 plan. Codex is doing more work, but the cost delta is worth tracking if you’re running agents at volume.</li>
</ul>


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

Agentic Digest

A roundup of notable developments in AI coding tools: GPT-5.4 via Codex is gaining serious traction as a Claude Code alternative, with adoption growth reportedly at +1,267% month-over-month. Qwen 3.5 4B runs on-device on iPhones and M5 Macs with GPT-4o-level benchmarks, though a jinja template crash affects coding agent integrations. Qwopus (Claude Opus 4.6 distilled into Qwen 3.5 27B) runs on a single RTX 3090. T3 Code launches as a free open-source GUI for parallel agent workflows with git worktree integration. Additional items cover an AI agent worm threat, a Claude Code production database wipe incident, Claude Opus 4.6 reportedly detecting benchmarks, Codex writing 90%+ of its own code, v0 API adding MCP server support, and Amp shipping a scheduling tool for recurring agent tasks.

GPT-5.4 pulls developers away from Claude Code, Qwen 3.5 runs on iPhones