Explore what's new in AI

JourneyChat lets multiple agents share memories in one group chat

Shared memory across agents is a useful primitive for delegation, coordination, and multi-agent debugging.

A Claude Code paper says the harness is most of the system

The decomposition argues permissions, memory, tools, and compaction dominate the product, not just the model call.

W&B iOS app adds live run logs and remote stop control

Mobile observability means you can inspect experiments and stop runaway jobs from your phone, not just your laptop.

Profiling skill ranks bottlenecks before optimization

It forces a baseline, hotspot table, and handoff artifact, helping agents stop guessing and start fixing the real bottleneck.

Qwen3.6-Max-Preview posts strong tool-use and coding benchmarks

The preview looks competitive on web, tools, and coding, making it worth a real side-by-side eval against frontier peers.

Nous Research ships Step 3.5 Flash to Portal with a free window

The free window makes it easy to compare a fast, cheap model inside a live portal, which is handy for routing and cost tests.

Cursor says plan mode helps cloud agents ship cleaner PRs

Have the agent plan first, then let overnight runs ship PRs. That reduces thrash when multiple agents are tackling one issue.

CopilotKit open-sources generative UI for agent apps

The demo turns model output into live components, showing one practical path from text generation to usable product UI.

HyperSkill turns websites into reusable agent skills

Feed it a URL and it turns the page into a callable skill, a useful pattern for turning web apps into agent capabilities.

discrawl 0.3.0 adds Git-backed sync for Discord archives

Git-backed archiving makes private Discord logs searchable and reproducible, which is useful for teams trying to preserve chat knowledge.

AgentFlow coevolves experts for task-specific skills

The paper trains specialized agent populations that beat a much larger single model, a strong argument for diversity over scale.

Ramp adds a controller after agents overshoot token spend forecasts

Ramp’s controller snapshots tasks, then asks a model for permission to spend more, a clean way to put guardrails around token burn.

Simon Willison shows how to force adaptive thinking in Claude Opus 4.7

The settings combo makes Opus think more deliberately before answering, useful when prompt hacks are too brittle.

Codex autocompaction lets headless agents run longer jobs

Automatic compaction keeps long agent sessions alive without a human babysitter, which matters once runs stretch into hours.

turbopuffer stem cache nearly doubles stemming throughput

Skipping repeated stems makes tokenization almost 2x faster in the hot path, a simple optimization with real serving impact.

Codex Chronicle uses screen history to improve agent memory on Mac

Screen-grounded memory can reduce repeated setup and help Codex pick up context faster after idle periods or app switches.

Browserbase benchmark cuts computer-use verifier errors to 1%

Browserbase says its new verifier slashes evaluation mistakes from 45% to 1%, a big win for browser-agent reliability.

DSPy adds optimization chaining and begins decoupling LiteLLM

Chaining lets one optimizer generate candidates and another tighten them up, which should make agent tuning more modular.

OpenAI leak shows Agents Studio for always-on ChatGPT agents

The UI points to templates, Slack, skills, and scheduling, a big step toward agents as managed workflows inside ChatGPT.

Google Research's ReasoningBank turns agent failures into memory

The framework stores both successes and misses as reusable memory items, which is a practical path to better long-horizon agents.

OpenCode adds Ling 2.6 Flash free for a limited time

A free frontier-ish model in the CLI makes it easy to test routing and throughput without changing your stack.

Cursor partners with SpaceX to train Composer on coding workflows

Cursor says SpaceX will help tune Composer, a sign it’s treating coding agents as enterprise systems, not just chat features.