<p><strong>TLDR:</strong> GPT-5.4 is turning heads as a coding agent, with several engineers calling it their new daily driver for the first time since early Sonnet days. Claude Code remains dominant but its 2x usage multiplier is frustrating users, pushing some to experiment with Codex and Qwen 3.5. Amazon’s AI-assisted production disasters are getting detailed post-mortems, and the “comprehension debt” problem is getting real attention from practitioners.</p>
<hr>
<h2>GPT-5.4 gaining ground as a coding agent</h2>
<p>A few engineers are calling GPT-5.4 at xHigh their preferred coding agent right now, which is notable given how long Claude models have dominated that conversation. The reported upside: it stays on task without constant nudging and handles more work per session. There’s a catch though — it burns tokens faster, and at least one developer flagged that “fast” mode silently drops reasoning to low for OAuth users without disclosing it, which is a legitimate transparency problem.</p>
<h2>Claude Code usage limits drawing real complaints</h2>
<p>The 2x usage multiplier Anthropic rolled out is hitting people hard. Google Antigravity (which runs Claude under the hood) is getting the same complaints — Pro account limits described as “borderline unusable” for anything serious. Some users are routing around it with Claude-Mem, an open-source plugin claiming 95% fewer tokens per session and 20x more tool calls before hitting context limits. Worth watching, though those numbers come from the project’s own README.</p>
<h2>Amazon’s AI production incidents keep compounding</h2>
<p>The full picture of Amazon’s AI coding tool failures is ugly: Kuro deleted a production environment during a bug fix (13 hours to recover), Amazon Q pushed bad code that wiped 120,000 orders, and a follow-on outage killed 6.3 million orders in a single day. Amazon’s response was to mandate senior engineer sign-off for AI-assisted code — while having already laid off 16,000 engineers. The structural tension there is obvious.</p>
<h2>Comprehension debt is the new technical debt</h2>
<p>An Anthropic study found engineers using AI assistance scored 17% lower on code comprehension tests than those who didn’t. The broader argument: teams are accumulating invisible risk as AI generates code faster than anyone can meaningfully understand it. Tests pass, velocity looks fine, but nobody can explain the design decisions. Unlike technical debt, comprehension debt doesn’t show up in any standard measurement until something breaks badly.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Superset (open source):</strong> New macOS terminal app for running Claude Code, Codex, and other CLI agents in parallel, each isolated in its own git worktree — addresses the context-switching overhead of multi-agent orchestration.</li>
<li><strong>Qwen 3.5 27B + Hermes:</strong> Getting positive mentions as a fast local combo for getting simple concepts running quickly, especially via Ollama.</li>
<li><strong>OpenClaw + Ollama:</strong> Ollama can now be configured as a provider in OpenClaw; the team recommends Kimi K2.5 on Ollama’s cloud with thinking disabled for near-instant tool calling.</li>
<li><strong>Dual-loop BDD for AI agents:</strong> Justin Searls published a practical <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> prompt for running dual-loop BDD with Claude Code — outer integration test loop, inner red-green-refactor — after finding simple TDD caused agents to chase coverage while missing integration correctness.</li>
<li><strong>Amazon’s AI governance response:</strong> Senior engineer sign-off is now required for AI-assisted code at Amazon, a policy that’s hard to square with the scale of recent layoffs.</li>
<li><strong>GoGogot:</strong> Lightweight self-hosted AI agent in Go, ~15MB binary, ~10MB RAM, deploys via single Docker command — positions itself as a no-telemetry OpenClaw alternative with 27 built-in tools.</li>
<li><strong>Codex + Notion NYC event:</strong> In-person event March 17 in New York featuring Codex demos and practical workflow sessions.</li>
<li><strong>GitHub Spark:</strong> GitHub’s AI micro-app builder is now available with Copilot Pro+ at $39/month; native repo context is the differentiator over Lovable and similar tools.</li>
<li><strong>VibePod CLI:</strong> Open-source tool for running any major coding agent in isolated Docker containers with zero config, plus a local dashboard for comparing token usage across agents.</li>
</ul>


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

Agentic Digest

GPT-5.4 is emerging as a competitive coding agent, with engineers praising its focus and session capacity, though token burn rate and a transparency issue with reasoning modes are noted drawbacks. Claude Code's 2x usage multiplier is frustrating power users, driving interest in workarounds like Claude-Mem and alternatives like Qwen 3.5. Amazon's AI coding tool failures are detailed — including a production wipe of 6.3 million orders — prompting a senior engineer sign-off policy that sits awkwardly alongside mass layoffs. An Anthropic study highlights 'comprehension debt': AI-generated code passing tests while engineers lose understanding of design decisions. Also covered: Superset terminal for parallel agent runs, GoGogot lightweight self-hosted agent, GitHub Spark availability, VibePod CLI for isolated agent containers, and dual-loop BDD workflows for Claude Code.

GPT-5.4 pulls ahead on coding, Claude Code hits usage limits