Explore what's new in AI

Mastra adds schema-validated request context and a NestJS module

Schema-validated tool context makes agent calls safer and easier to reason about, while the NestJS module lowers the friction of wiring it into apps.

Replit launches Parallel Agents, letting teams run up to 10 agents on separate app copies

Each agent gets its own isolated copy of the app, which makes side-by-side exploration practical. It is a strong pattern for faster multi-branch agent work.

Google updates Gemini docs with an Interactions API built for agents and tools

The docs now expose an agent-optimized API surface alongside the older generateContent path. That makes the platform easier to wire into tool-using systems.

Jimmy Lin says AI-agent reproducibility needs artifacts built for machines

The paper argues human-oriented instructions are not enough, agents need self-contained artifacts they can rerun. That is a strong framing for reproducible ML work.

RLance Martin explains Claude Managed Agents' grader loops and dreaming

The post breaks down self-verification and self-learning as reusable agent patterns, not vague magic. That is useful for long-running workflows with rubrics and memory.

Recursive Agent Optimization trains agents to delegate to other agents

The paper frames delegation itself as a learning problem, which is exactly where multi-agent systems get interesting. It is a promising path for hierarchical workflows.

Simon Willison finds multiple Claude Code processes eating about 30 GB on Mac

A few agent sessions can quietly dominate laptop memory. Teams running lots of terminals should watch background agent footprint, not just token spend.

OpenClaw shows an end-to-end test for one OpenAI-compatible agent powering another

The test proves tool calls, streaming, and role:tool handoff survive a nested agent loop. That's a smart compatibility check for OpenAI-style backends.

The Memory Curse paper says longer histories can reduce cooperation in agents

Across seven models and four games, expanding accessible history hurt cooperation in most settings, and a LoRA fix restored some of it. Bigger context is not always better.

PACT benchmarks AI negotiation in a 20-round buyer-seller bargaining game

The setup makes bargaining measurable by forcing message exchange, bids, and asks over thousands of matchups. That is useful if you care about agent strategy, not just QA.

AgentPlug uses a USB-C dummy plug to keep Macs awake in clamshell mode

It is a low-tech but practical way to keep background agents running without brittle keep-awake software. The tradeoff is simple physical control instead of a process.

Simon Willison runs LLM CLI tools from shebang lines with YAML templates

The trick turns prompts into executable scripts, which is handy for repeatable little automations. It also shows a simple pattern for bundling tools and functions.

Trimmy adds Claude Code prompt trimming and hides its menu bar icon

The utility now cleans up Claude Code prompts before paste, which reduces friction when moving agent context between apps. Small DX tools like this matter daily.

Codex rewrites uv into Bun and TypeScript, then passes 1,134 tests

This is a strong end-to-end demo: a full language migration, parity checks, and a green test suite. It is exactly the kind of artifact that makes agent claims credible.

Codex leak points to a new Ultrafast tier for latency-sensitive work

The config diff shows OpenAI experimenting with a tighter speed tier alongside Fast and Priority. That hints at more explicit latency tradeoffs in agent UX.

Cursor brings agent task delegation into Microsoft Teams

Mentioning Cursor in a channel can now spin work into an agent and return a PR for review. It is a clean example of chat surfaces becoming control planes for code.

OpenAI's realtime meeting assistant turns standups into ticket movement

The repo shows a voice-to-action workflow where speech updates a ticket board automatically. That is a practical pattern for teams exploring real-time agents beyond chat.

Claude Cowork books 8 flights and 5 hotels in one shot

This is a rare real-world browser agent report with a clear outcome and no babysitting. It suggests the agent layer is getting reliable enough for high-friction workflows.

xAI leaks Grok Build, a desktop coding app for macOS, Windows, and Linux

The preview points to xAI joining the desktop coding-agent race with a native app instead of a web wrapper. That matters if they ship local workflow and plugin hooks.

Gemini adds travel planning across Gmail, Photos, Search, and YouTube history

Google is wiring personal context into itinerary building, not just chat. That points to assistants that can pull from connected apps and do useful planning.

Google's Gemini Omni leaks show a video model with editing and watermark removal

Screenshots and demos point to a new video model with chat-based edits, object replacement, and short clips. If real, Gemini is moving toward a full media editor.

OpenAI adds a Developers plugin to Codex for API scaffolding

Codex can now generate app and agent code against OpenAI APIs from inside the editor. That pushes it toward an opinionated API-building workflow.

Birdclaw turns your Twitter archive into a searchable memory for Codex

The tool lets you query old likes and bookmarks instead of hunting through exports by hand. That is the kind of personal context layer agents will increasingly need.