<p><strong>TLDR:</strong> Anthropic quietly doubled Claude usage limits for off-peak hours and weekends through March 27, a move that landed well with the Claude Code crowd. OpenAI’s Codex is logging serious autonomous sessions — one developer reported a single 10-hour run on a complex economics research task. The MCP vs. CLI debate keeps grinding on, with a new lightweight HTTP protocol called HALO entering the mix. Meanwhile, Qwen’s 3.5-Flash model is getting attention for near-frontier performance without the infrastructure overhead.</p>
<hr>
<h2>Anthropic doubles Claude usage limits through March 27</h2>
<p>The boost applies across all plans — Free through Team — during weekends and non-peak weekdays (outside 5–11am PT). Anthropic framed it as a “Spring Break” promotion and tied it to a thank-you to the Claude Code community. One developer noted the timing coincides with the Claude Opus 4.6 drop and some recent outage frustration. Whether this is goodwill or capacity management, the practical effect is the same: more headroom for long agentic sessions.</p>
<h2>Codex running 10-hour autonomous coding sessions</h2>
<p>A developer reported a personal record: just under 10 hours in a single Codex session using GPT-5.4 on a complicated economic research prompt. No flashy app, just a hard problem and an agent grinding through it. Separately, another developer noted that after Claude Code spent 8–10 cumulative hours failing to solve a specific problem, Codex solved it quickly. Neither data point is a benchmark, but the pattern is worth watching.</p>
<h2>CLI-Anything makes any software agent-controllable</h2>
<p>CLI-Anything is an open-source tool that auto-generates CLI interfaces for arbitrary software — GIMP, Blender, LibreOffice, OBS — so AI agents can drive them. It runs a 7-phase pipeline (analyze, design, implement, plan tests, write tests, document, publish) and ships as a Claude Code plugin with OpenCode and Codex support. The project has 1,508 passing tests across 11 validated applications. Generated CLIs support REPL mode, subcommand scripting, undo/redo, and structured JSON output for agent consumption.</p>
<h2>HALO proposes self-describing HTTP APIs with no MCP layer</h2>
<p>HALO (HTTP API Language for Operations) lets any HTTP API expose a tool manifest via the standard OPTIONS verb and a custom Accept header — no MCP server, no registry, no proxy. A FastAPI package handles server-side registration and provides adapters for Microsoft Agent Framework and Semantic Kernel. Key properties: zero schema drift (schemas derived from Pydantic models), auth-aware scoping, lazy schema loading, and direct agent-to-API calls. It’s a minimal-footprint alternative for teams that find MCP overhead hard to justify for internal APIs.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Qwen 3.5-Flash architecture:</strong> Uses linear attention and sparse MoE, targeting near-frontier performance without data-center-scale compute — getting real traction among developers who want capable models they can actually run.</li>
<li><strong>Ramp shipped 500+ features with 25 PMs:</strong> The Claude Code workflow they used has three phases; details are sparse but the ratio is striking enough to pay attention to.</li>
<li><strong>Anthropic A/B testing plan mode silently:</strong> A $200/month Claude Code user found a system-level instruction hard-capping plans at 40 lines with no opt-out. The post argues professional dev tools need transparency, not consumer-app-style experiments.</li>
<li><strong>Context Gateway (YC-backed Compresr):</strong> An open-source agentic proxy that pre-computes conversation summaries in the background so context compaction is instant when limits hit — works with Claude Code, Cursor, and others.</li>
<li><strong>OpenClaw beta adds live browser control:</strong> Uses Google’s remote debugging in the latest Chrome (chrome://inspect#remote-debugging), with the agent deciding when to invoke it or accepting manual direction.</li>
<li><strong>Gemini anti-persona update causing issues:</strong> Multiple users reporting increased hallucinations and unwanted persona triggers after a recent Gemini update, with at least one threatening to cancel their subscription.</li>
<li><strong>Windsurf credit costs:</strong> Users reporting $0.30 per request on Windsurf; the practical workaround is batching tasks into longer prompts and routing simpler requests to Sonnet or Haiku.</li>
<li><strong>GPT-4 turns 3:</strong> The sketch-to-website demo that made the rounds in 2023 is now a historical artifact. We’re three years into that shift and the pace hasn’t slowed.</li>
</ul>


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

Agentic Digest

Anthropic temporarily doubled Claude usage limits for off-peak hours and weekends through March 27 as a Spring Break promotion for the Claude Code community. OpenAI's Codex is demonstrating extended autonomous sessions, with one developer reporting a nearly 10-hour solo run on a complex economics research task. New tooling is emerging around AI agent control: CLI-Anything auto-generates CLI interfaces for arbitrary software so agents can drive tools like GIMP and Blender, while HALO proposes a lightweight HTTP-native alternative to MCP for exposing APIs to agents. Other notable items include Qwen 3.5-Flash's efficient architecture gaining traction, Ramp shipping 500+ features with 25 PMs using Claude Code, and reports of Gemini hallucinations after a recent update.

Anthropic doubles Claude usage limits, Codex hits 10-hour solo sessions