#Agent

AIReplit says its Desktop app for Windows is in private preview with Microsoft and NVIDIA, building and running apps in isolated sandboxes on a user's PC. The company also lets users work across projects from one chat, including finding projects, reading their files, and sending them tasks. A new TikTok Ads MCP lets users create, launch, and track TikTok ads from Replit.

AIMicrosoft researchers introduce CABRA, a framework that generates synthetic coding tasks with one difficulty dimension varied at a time. Across 6,840 tasks, plain LLMs degraded as tasks grew, while agents stayed near-perfect by offloading work to tools such as grep. On SWE-bench Verified, counts of reading and analysis calls correlated with agent failures at -0.200, versus -0.159 for lines edited.

AIPrime Intellect says Prime Agent used a swarm of over 2,000 agents to rewrite itself end to end in Rust over two weeks. The rewrite ran across 10,000+ sandboxes and over 200 billion GLM-5.3 tokens, and the company says usable input now arrives about 13 times faster with 83% less startup memory.

AIArena says Agent Arena measures AI agent capabilities using causal tracing instead of pairwise preference voting. The method tracks five signals from real-world, long-horizon agentic sessions, according to the post, which points to a thread for more details.

AITinker says it made efficiency improvements to support scaling long-context reinforcement learning and is passing them on through price cuts of up to 70%. GLM-5.3-Flash and DeepSeek-v4.1-Flash are now live on the platform for cost-efficient long-context work.

AIPrime Intellect says its Prime Agent orchestrated over 2,000 agents over two weeks to rewrite the agent in Rust. The run used more than 10,000 sandboxes, over 200B GLM-5.3 tokens, and 16,000 agent-to-agent messages. The company says the rewritten agent reaches usable input about 13 times faster and uses 83% less startup memory.

AIPrime Intellect says it preserved its TypeScript version's features and behavior by giving agents objective parity checks. The checks diff terminal frames, compare session transcripts and model requests, check daemon protocol messages, and audit every feature. Agents could see where behavior diverged and fix it before changes merged.

AIA root agent splits a rewrite into dependent tasks, with each task run through a Planner, Implementer, Reviewer, and Verifier state machine. Each implementation must pass independent review and verification in a fresh Prime Sandbox before merging, and failed checks send the task back to the implementer. Tasks can run in parallel without skipping these checks.

AIPrime Intellect says Prime Agent, after reaching feature parity, ran its own runtime benchmark suite and tested candidate changes against the current build. Changes that passed parity checks and independent review became the baseline for the next experiment. The loop moved work off the startup and render paths and released memory after large sessions loaded.

AIElevenLabs is releasing synthetic voice detection in ElevenAgents to help businesses identify AI agents calling on behalf of individuals, companies, or bad actors. The company is also joining the Personal Agent Protocol working group to help define how agents interact.

AIArena says Agent Arena replaces pairwise preference voting with a causal tracing methodology. The method measures five signals from real-world, long-horizon agentic sessions.

AIVercel says its CLI now lets AI agents purchase domains directly. The post links to a Vercel changelog entry with details.

AILangChain says LangSmith LLM Gateway now supports the OpenAI Decisions API for low-latency agent inference. The gateway provides centralized controls for model fallbacks, data redaction policies, and spend limits.

AIPrime Intellect says inference scaling works, but agents eventually run into the same problem: their context windows fill up. The company argues in an essay that the only scalable fix leads to swarms of agents.

AIPrime Intellect says compaction summarizes a full context window and passes the summary to the next one, but each summary is a guess about what will matter later. Offloading memory to a filesystem or REPL avoids that guess, but files cannot reason, so the agent must load them back into its window and spend the context it was trying to save.

AIPrime Intellect says agents should communicate directly with each other, turning a tree structure into a mesh. Context is distributed across the swarm, so no single context window must hold all information, and agents specialize over time.

AIEric Zakariasson shares four Grok @bot templates for building on X: launching, threat hunting, threat intel, and X API integration. He says all four can be used today.

AIElvis Saravia says Codex's new composer predictions match a tool he has run for months in his agent orchestrator. He says his version is tunable, adapts to his preferences, and uses smaller models such as Haiku and Luna. He calls it a quality-of-life feature that makes agents more proactive.

AIOpenAI says its dot can start work in Codex and follow up on existing threads, drawing on ChatGPT conversations, Codex threads, and automations. The dot also decides whether to continue a thread or start a fresh one, and can review and edit Scheduled Tasks in ChatGPT Work.

AISpaceXAI says Grok Bot users can ask their bot to claim its own email address for contacting others and signing up for newsletters. Testing Catalog reports the bot subscribed it to its daily AI Brief newsletter without issue. Admins must enable the feature for their team, and it is rolling out to users starting today.

AIAnthropic's Lydia Hallie clarifies that Claude Code's auto-compact replaces the whole conversation with a short summary. On 1M-context models it triggers around 967K tokens, and each message before that point re-reads the full conversation, mostly from cache. Running /autocompact 400k makes compaction trigger at 400K instead.

AIOpenAI says composer predictions in Codex is now in beta for Pro users. The feature suggests a user's next message based on their conversation and how they phrase requests. OpenAI calls it one of the most loved features its team has tested internally.

AIAnthropic's ClaudeDevs account says it has let in every Pro and Max user from the Claude Code Projects waitlist. The post links a 4-minute walkthrough video for new users getting started with the feature.

Why it matters: The post shows Claude Code Projects access opening to Pro and Max users from the waitlist, with a walkthrough for new users getting started.