Two things moved on the same day. Claude Code made auto mode the default permission mode, Hermes shipped a self-pacing loop command, and a run of AI Engineer talks argued that reliable browser agents are an architecture problem teams can solve now rather than a wait on better models. At the same time the verification layer thickened: GitHub published a stacking workflow for reviewing thousand-line agent diffs, a repair study measured how bloated agent patches are even when the tests pass, Cloudflare shipped detection for MCP traffic, and Anthropic answered the practical questions about output watermarking. The autonomy is arriving faster than the review habits around it, which is where the next month of work sits.
Auto mode becomes the default permission mode in Claude Code for Pro, Max and Team. Existing explicit defaults are preserved, Shift+Tab still switches, and defaultMode pins one in settings. The companion setup command scans the repository and proposes trusted repos and domains for review, so the classifier's accuracy depends on how much of the environment it has been shown.
Open weights had a loud day. GLM-5.3 posted frontier-adjacent agentic coding scores at roughly 750B parameters, a third of Kimi K3, with the gains coming from extended post-training on the GLM-5.2 base. Qwen3.8-27B landed on Ollama the same day, and hands-on testing across browser workflows, C++ game code, CAD and multimodal tasks gives a task-level read on what that size can carry locally.
A dated brief from the vibe-coding frontier. Today’s Intel.
If post-training alone produced the GLM-5.3 jump, the distance between American and Chinese open models is a process gap, not a compute gap. That is the sharper version of the argument now running, and it changes what a catch-up plan should even budget for.
Review is the bottleneck once an agent can write 1,700 lines in one pass. GitHub's answer is stacking: one concern per layer, CI evaluating each layer against the stack base, read top-down and review bottom-up, with the gh stack CLI handling rebase and sync as feedback lands. A case study on dismantling an architectural invariant across 189 files in a 717k-line codebase reaches the same place from the other direction, with specification-first convergence standing in for the human review that was not available.
A passing test is not a proxy for a clean patch. New work on LLM program repair quantifies how consistently agent diffs are bloated beyond the fix and finds that asking for minimality does not correct it, which puts a measurable cost on accepting green CI as the whole review. A separate study on generated webpages finds the same blind spot in frontend evaluation, where visual fidelity is usually scored in one fixed browser and cross-environment rendering is never checked.
Controls for agent traffic arrived from the network side. Cloudflare can now detect MCP traffic, which surfaces shadow servers and lets a team force agents through an approved path, and internal applications built with coding agents can be made private by default through an account-level policy rather than per-builder auth wiring.
Anthropic published an FAQ on text watermarking: implemented for EU AI Act compliance under a Code of Practice other labs have signed, no hidden characters, no extra tokens, no price change, and no trace back to a person, organization or chat. Commentary is already circulating about what the marker means for shipped code and what the countermeasures are, so treat detection as a weak signal in both directions until the mechanism is better understood.
Browser agents fail on real sites for reasons that are architectural, not model-limited. The AI Engineer talks name the specific splits that separate production systems from demos, including escalation strategy and where control sits between agent and harness, while a computer-use talk argues that many published benchmark numbers are artifacts of deterministic environments and undersized confidence intervals.