|
Issue · November 20, 2026
Gemini brings computer use to the Mac, while production agent workflows and reported privacy failures show why useful automation needs explicit permissions and independent checks.
AI in general
Frontier models, research and policy
|
3 stories |
Score · 95 / 100
Gemini gains computer use on Mac, with a separate background workspace
Source: Google — November 19, 2026
COMPUTER USE · PRODUCTIVITY · AGENTS
Gemini’s Mac app can now click, type, navigate menus, and fill forms across desktop apps, local files, and websites. It operates in an isolated background window with a live preview, leaving your main screen and cursor available. The practical advance is cross-application execution—not just advice or generated content—although Google’s examples are demonstrations rather than independent reliability tests.
What matters
- You choose which apps it can access, can pause execution, and must approve key actions.
- Examples include extracting PDF details into shipping forms, building slides from dashboards, and batch-processing creative assets.
- Requires Apple Silicon, macOS 15.0+, an adult personal account, and Google AI Plus, Pro, or Ultra. The English rollout excludes the EEA, UK, Switzerland, and Nigeria.
VerdictSKIM** — Check eligibility and the workflow examples; test first on reversible, low-stakes tasks.
Score · 92 / 100
Meta’s Muse reportedly mixes conversations across accounts and surfaces private information
Source: TechCrunch, citing Reuters — November 19, 2026
PRIVACY · AGENT SAFETY · META
Meta employees reported that Muse surfaced messages and information that did not belong to their conversations, according to Reuters reporting on internal company posts. Examples included unfamiliar bank balances and a reminder about a Signal message that an employee said she had not authorized Muse to access. These are reported incidents, not a publicly established technical diagnosis; TechCrunch says Meta had not responded to its request for comment.
What matters
- The central concern is account and permission isolation, not merely inaccurate answers.
- The article does not establish whether each example reflects real cross-user disclosure, fabricated information, or another failure mechanism.
- For personal agents, treat access to messaging, financial information, and persistent memory as separate permissions—not one blanket authorization.
VerdictSKIM** — The examples matter; the underlying cause and extent remain unresolved.
Score · 86 / 100
ChatGPT adds live dictation and removes its 10-minute recording limit
Source: The Verge — November 19, 2026
CHATGPT · VOICE · UX
ChatGPT can now transcribe microphone input while you speak instead of waiting until recording ends. You can reveal the live transcript by selecting the waveform, then edit the resulting prompt before submitting it. The feature is rolling out globally across all plans on web, desktop, iOS, and Android.
What matters
- OpenAI is also removing the previous 10-minute limit for prompt-box dictation.
- This is distinct from two-way voice conversations and the separate, plan-limited meeting-recording feature.
- The immediate benefit is faster correction of names, technical terms, and long spoken prompts.
VerdictSKIP** — You have the useful details; look for the feature in the app.
Software engineering
Coding agents, developer tools and infrastructure
|
3 stories |
Score · 95 / 100
OpenAI’s Performance Factory turns alerts into investigated, reviewed, and verified fixes
Source: OpenAI — November 19, 2026
CODING AGENTS · RELIABILITY · PRODUCTION ENGINEERING
OpenAI describes an internal Codex-powered system that investigates performance alerts, proposes fixes for engineer review, and checks recovery after deployment. In just over three months, its regression workflow tracked more than 1,700 investigations and identified root causes in over 600 cases. The transferable lesson is the architecture: durable task state, explicit workflow transitions, event-driven resumption, and a separate agent tasked with challenging diagnoses.
What matters
- Agents use the team’s existing metrics, traces, profilers, and code access through scoped service identities.
- Early agents sometimes accepted the first plausible explanation; adversarial review now tests counterfactuals before accepting a diagnosis.
- Root-cause counts are not counts of successfully deployed fixes, and the results are OpenAI’s own operational report.
- Human review remains a bottleneck; fully autonomous merging is described as something being explored.
VerdictREAD FULL** — A concrete production case study worth borrowing from if you build agent workflows.
Score · 93 / 100
Gemini API Tool Retrieval loads function schemas only when needed
Source: Google — November 19, 2026
API · TOOL CALLING · CONTEXT EFFICIENCY
Google’s preview feature keeps large tool libraries available without placing every parameter schema in the initial context. In server-side mode, Gemini sees deferred tools’ names and descriptions, requests the relevant schemas, and produces the function call within the same API request. This targets a real scaling problem for agents connected to dozens or hundreds of functions.
What matters
- Add
tool_search and mark infrequently used tools with defer_loading: true; remote MCP servers can also be deferred.
- Available in the Interactions API for Gemini 3.6 and later, including Gemini 4 models.
- There is no separate retrieval fee, but loaded schemas and retrieval overhead still consume billed tokens.
- Client-side retrieval supports custom registries and permission filtering, at the cost of another round trip per search.
VerdictREAD FULL** — Useful implementation detail if tool definitions are inflating your prompts; measure savings on your actual traffic.
Score · 85 / 100
AgentCore adds deterministic lifecycle hooks and stronger managed harness defaults
Source: AWS — November 19, 2026
AGENT INFRASTRUCTURE · GOVERNANCE · AWS
AWS details an AgentCore harness upgrade built on the open-source Strands harness, adding context management, planning, delegation, and lifecycle hooks. Synchronous Lambda hooks can allow or deny consequential operations at defined checkpoints rather than relying only on model instructions. This is a newly published technical explanation of a rollout that began November 11—not a claim that every capability launched yesterday.
What matters
- A
before_tool_call hook can block a proposed action while letting the agent continue, for example to explain an escalation.
- SNS and EventBridge hooks are notifications; they do not block execution. A hook after completion cannot retract already streamed output.
- Omitted settings follow evolving managed defaults, including the recommended model and tool set; explicitly pin configuration when your security boundary must remain fixed.
VerdictREAD FULL** — Especially for existing AWS users; the hook semantics and changing defaults deserve attention.
Design & creative
Creative workflows and user experience
|
1 story |
Only 1 strong recent item found.
Score · 84 / 100
Cosmos connects visual generation, labeling, and evaluation into repeatable dataset workflows
Source: NVIDIA — November 18, 2026
VIDEO GENERATION · SYNTHETIC DATA · CREATIVE PIPELINES
NVIDIA demonstrates agent-orchestrated workflows for generating and transforming images and video: changing weather or lighting, creating rare events, and adding defects to clean inspection images. Cosmos Data Factory combines five reference blueprints with six reusable skills covering curation, generation, simulation, labeling, and orchestration. This is most relevant to technical visual-production teams and physical-AI developers, rather than everyday graphic design.
What matters
- Video augmentation aims to preserve scene structure while changing conditions; automated evaluation checks requested changes, artifacts, and physical plausibility.
- Generated outputs can include annotations and provenance manifests, making them more useful downstream than isolated media files.
- “Two prompts” describes orchestration after environment preparation—not a lightweight, infrastructure-free creative app. NVIDIA’s cost examples cover compute only, excluding setup, storage, agent use, and human review.
VerdictSKIM** — Read deeply only if you need controlled visual variants or training-data pipelines; validate quality against real-world examples.
Open-source watch
Projects gaining meaningful traction
|
3 stories |
These are current discovery picks, not necessarily newly created repositories. Traction figures come from the supplied GitHub Trending snapshot.
Project · 01
Diagram Design — cathrynlavery/diagram-design
DESIGN · DEVTOOL · AGENT
An agent skill for producing editorial diagrams as self-contained HTML and SVG, with optional brand styling. It is interesting because the output remains inspectable and editable rather than becoming a flattened generated image.
What matters
- The snapshot reports 402 stars today; the current README lists 51 diagram types, ahead of the snapshot’s older description.
- Supports multiple agent hosts and use cases such as architecture diagrams, sequences, and prioritization quadrants.
- Review skill permissions and network behavior before installation; manually check generated layouts and factual relationships.
VerdictTRY** — A bounded experiment for your next architecture explanation or presentation diagram.
Project · 02
Stitch Skills — google-labs-code/stitch-skills
DESIGN · AGENT · DEVTOOL
A collection of Agent Skills and plugins for working with Google Stitch through its MCP integration. It offers a reusable bridge between design workflows and coding agents instead of requiring each agent to rediscover the same process.
What matters
- The snapshot reports 303 stars today.
- Uses the Agent Skills standard and supports several coding-agent environments; it depends on the Stitch integration rather than functioning as a standalone design tool.
- Apache-2.0 licensed, but the repository explicitly says it is not an officially supported Google product.
VerdictTRY** — Worth testing if you already use Stitch and want a repeatable design-to-build workflow.
Project · 03
Onlook — onlook-dev/onlook
DESIGN · FRONTEND · AI
Onlook combines a visual canvas with a real codebase so designers and developers can work closer to the implementation. Its November 19 Launch HN post describes a desktop workflow that works with the coding agent you already use; the repository independently confirms its visual code-editing focus and Apache-2.0 license.
What matters
- The repository shows roughly 27,000 stars; popularity is an adoption signal, not proof of production fit.
- Particularly relevant to teams trying to reduce the gap between mockups and frontend changes.
- Test on a branch and inspect the diff before adopting it for an established design system; compatibility with your stack is the deciding factor.
VerdictWATCH** — Promising for designer–developer collaboration, but trial it on a small real component before changing your workflow.
Editor’s note
Fresh, actionable developments in desktop automation, production agents, and editable visual workflows—prioritizing explicit controls and concrete implementation details over repeat model-launch coverage.
30-second feedback
How useful was today’s digest?
Your choice opens the short form with your rating filled in. Or share a quick note.
Prefer reading on Telegram? Subscribe to The Merpati Post →
The Merpati Post · Daily AI Briefing
|