Explore what's new in AI

Google Flow Agent grounds image and video generation in Maps Street View

Grounding scenes in real locations should improve accuracy and make location-aware media generation less generic. That matters for workflows tied to places.

Tetsuwan turns wet-lab work into cloud lab workflows

The system wraps biology tools and research databases into a workflow agents can orchestrate, pushing lab automation closer to normal software.

cwolferesearch defines the agent loop around model, tools, context, and environment

The value of this breakdown is making orchestration and termination explicit, which helps teams reason about what the harness is doing.

HoloDesktop CLI lets agents operate a local desktop

The CLI brings computer use onto your own machine with low latency and full privacy, and plugs into harnesses people already use.

DynamicWebPaige's Chrome extension records browser actions into skills

Recording real browser behavior is a practical bridge from UI interaction to reusable agent instructions, instead of hand-writing brittle steps.

Cursor in Notion lets specs turn directly into tasks and PRs

Notion's Cursor integration collapses one more handoff between docs, implementation, and review, which is exactly where teams lose time.

Manus adds autoscale and reserved hosting modes for published apps

Teams can now choose idle cost versus always-on execution explicitly, which is a useful deployment knob for agent-built apps.

Hermes Agent's /learn skill turns docs into a learning plan

The skill creates a learning plan and an artifact hub, which makes research sessions repeatable instead of starting from scratch every time.

Sakana AI brings Fugu-Ultra to OpenRouter as a multi-model system

Sakana is leaning into orchestration over a single monolith, which makes the system easier to test, route, and swap as models evolve.

Agent-native memory paper breaks memory into storage, retrieval, and maintenance

The paper evaluates memory modules separately instead of treating memory as one black box, which is the right way to reason about cost and failure modes.

Perplexity adds Computer for Counsel for legal research workflows

Perplexity is wiring research sources and matter tools into one legal workflow, which makes it easier to pull citations without bouncing across apps.

Eric Xing's Critique of the Agent Model defines agency across five dimensions

The paper separates goal, identity, decision-making, self-regulation, and learning, giving teams a cleaner way to argue about what actually counts as an agent.

OpenAI updates GPT-5.5 Instant to follow intent and constraints better

OpenAI says the model now responds more naturally to intent and handles complex constraints more reliably, which should reduce brittle assistant behavior.

Exa Connect adds agent access to ZoomInfo, Crunchbase, and Similarweb

Exa is turning paid provider data into an agent-accessible layer beyond the public web, which could beat scraping for reliability and freshness.

Google AI Studio says users built 1M Android apps

Google says AI Studio crossed 1M Android apps after native app building launched in May, which is a strong sign the workflow is finding real adoption.

Nous Research adds a creative-ideation skill that routes prompts through 22 methods

The skill gives agents a structured way to move out of stale first-pass thinking, which is useful when you want novelty without losing feasibility.

Matt Pocock says no-op phrases bloat agent skills without changing output

He argues that filler instructions usually do nothing, so removing them makes skills cheaper, easier to test, and easier to maintain.

Matt Pocock's /loop-me skill interviews your work to find AI delegation opportunities

The skill turns a vague workday into concrete automation opportunities, which is a pragmatic way to push more routine tasks onto agents.