Good morning, and welcome back to Techzip AI Newsletter.

Today’s Techzip covers how Claude searched roughly 200,000 reverse transcriptases and surfaced a new enzyme system, why Meta Charm puts Muse on a handheld device, how Cursor reports a 7% cut in production token costs, and how an OpenAI agent breached Australia’s health-data portal. We also look at Google Cloud agents scanning and patching code before production.

You just shipped a new feature. But when customers do not know how to use it, support gets more tickets, sales has to demo the same flows again, and the content team has to remake guides whenever the product changes.

Vidocu's AI Recorder breaks that cycle. Describe the workflow, and its agent opens a real browser, clicks, types, and performs the steps itself. One run produces a walkthrough video, step-by-step screenshots, an article, and a quiz.

How it works:

Because the agent performs every click, Vidocu knows exactly which controls to zoom in on or highlight. You do not need to walk through the flow manually, record your screen, then start creating documentation.

AI Recorder is only the beginning. Vidocu also includes Studio, Course Builder, training quizzes, and Remix to turn webinars or recorded calls into short-form reels, blog posts, LinkedIn posts, and X threads.

Use VID50FM to Get 50% Off your first month to unlock AI Recorder.

Anthropic says Claude helped researchers identify a new enzyme system with CRISPR-like repeats. Claude searched scientific literature and genomic data, proposed experiments, and continued following signals that emerged during the search.

Anthropic says the process involved roughly 950 Claude sessions over 21 hours, a search across about 200,000 reverse transcriptases, and 19 reports. One agent also moved beyond the original search target to identify a repeat array in the enzyme system. That is the clearest sign of Claude's role here: processing a large hypothesis space and surfacing a direction outside the initial target.

Claude handled the search and candidate generation; scientists chose the research direction and performed the experiments. The enzyme's function, biological significance and mechanism remain unknown, and Anthropic presents this as an early research finding.

Meta introduced Meta Charm, a handheld device for accessing its Muse AI assistant. Announced Muse capabilities include avatar video calls, background tasks, Mac computer control, email support and commerce connectors.

Meta says voice access on its glasses is coming in the next few months. Its Realtime Avatar latency, resolution and capacity figures are self-reported, and not every avatar shown in the demos is available in the Muse app.

Muse now spans several surfaces: avatar communication, background tasks, computer control, email, commerce and glasses. Those capabilities are rolling out on different timelines.

Cursor says its engineering changes cut agent token costs by 7% without reducing quality in its production traffic. The technical post focuses on the operating layer around the model: how prompts are assembled, which tools are loaded, which files are read, and which context can be reused.

The changes include tighter prompts, selective tool loading, better cache handling and more compressed file reads. In long sessions, Cursor also counts the tokens generated by subagent coordination instead of looking only at the cost of each model tool call.

Cursor treats token efficiency as a whole-agent-harness problem, not just a model problem. The post focuses on five cost layers in a multi-step agent run: prompts, tool loading, caching, file reads and subagent coordination.

An agent can spend tokens not only on the model, but also on prompts, tool payloads, request assembly, caching, context compression, retrieval and coordination between subagents. The audit prompt from Eric Zakariasson (formerly at Cursor), traces each layer in that chain.

Use the prompt as a checklist for one specific agent workload, starting with a representative run. Ask it to inspect:

Once the costs are broken down, ask the prompt to rank the largest sources of waste and propose changes in priority order. Run the same workload again, then compare tokens, cost and output quality before and after to catch quality regression. The workflow helps locate the harness bottleneck before optimization.

The paper tests whether models can program agent artifacts and run them on held-out tasks. Results show a gap against human-built baselines, alongside adversarial behavior when systems are optimized against a target. The important shift is that the benchmark measures whether a model can build and adapt a workflow, not only complete a fixed task.

The paper models orbital AI through constraints on compute, connectivity, energy and thermal headroom across spacecraft, orbital networks, ground stations and cloud backends. When runtime is bounded by location, energy and connectivity, those physical constraints become part of the system architecture.

The interview recap focuses on context curation, durable guardrails and evaluations, deterministic tools, iteration limits and operational safety patterns. For teams building agents, the practical lesson is to narrow the action space, make tools inspectable and put explicit limits around loops.

Jeffrey Katzenberg argues that as AI expands content production, consent, compensation and recognition for creative work need to be treated as part of the system rather than handled afterward. It is a normative view of how creative industries might allocate rights and value as AI enters more of the production process.