#Agent

AIMicrosoft and Hugging Face released ThinkingBox, a benchmark that grades AI agents on the terminal backend state and side effects they leave behind rather than their final responses. Each of 507 stateful business tasks runs 20 times from a clean backend, and the post reports pass@1, pass@20, and observed 20/20 counts, plus cost per successful and per dependable task across 18 models. The harness and dataset are available on Hugging Face, with the OpenEnv interface for running evaluations.

Why it matters: The post shows why checking the database state, not tool calls or final replies, exposes agent failures, and gives a repeat-run method for judging reliability.

AIReplit CEO Amjad Masad argues that general models could train smaller, domain-specific successors on the fly when they detect a limited use case. He compares this to a just-in-time compiler that emits optimized code during execution. He says such specialized models could be cheaper, less vulnerable to prompt injection, and less harmful than general agents.

AIHarrison Chase, LangChain's leader, says his team is debating whether to build specialized agents or one mega agent. The post gives no conclusion or supporting detail.

AIPerplexity's Computer can now generate inline visualizations when users ask it to "Visualize" a topic, producing interactive widgets and animations within the thread. The feature is best used on Standard or High effort, and an example given is an inline 3D cutaway of a jet engine.

AIAmjad Masad of Replit and Alex Atallah of OpenRouter discuss why AI independence and model diversification matter for enterprises. They argue that depending on a single lab risks lock-in and that specialized agents may outperform one general superagent. The post presents the conversation as a podcast episode, the first Atallah has done since Stripe acquired OpenRouter.

AIYuchen Jin argues that the terminal is the wrong interface for coding agents, since managing many tabs creates cognitive overhead while context should persist. He says he rarely needs an IDE like Cursor because he seldom navigates the whole codebase now, calling the agent rather than the file the new primitive. He names the Codex desktop app as the best agentic UI for now, while noting the space is still early.

AIA consultant helping several companies adopt AI in engineering workflows says teams become much more productive and ship better software faster once they ramp up. The shift he recommends is from prioritizing human-maintainable code to building strong processes that validate what agents do, and he rejects the view that such software will later prove worthless.

AICMU researchers released SMDD-Bench, a benchmark of 502 small-molecule drug design tasks that use RDKit, ADMET-AI, and Boltz-2 as feedback loops. The authors argue that long-horizon planning, exploration, and learning from imperfect feedback remain open problems beyond math and coding, and the benchmark is available in Prime Intellect's Environments Hub for training with prime-rl.

AIPrime Intellect has launched Prime Inference, serving GLM-5.3 with vLLM at scale for agent workloads. The vLLM project credits the team's work on prefill/decode topology, scheduler bubbles, and reliable tool calls.

AIIndependent analyst Mostly Borrowed Ideas said he sold his Airbnb stake and added to Meta after testing Meta's Muse AI agent for about 10 days. He said Muse browsed Airbnb like a human, then found a farmhouse stay about 60% cheaper by booking directly with the host, suggesting AI agents could bypass booking platforms. He acknowledged Muse is slow, with a five-hotel price comparison taking 14 minutes.

AIPrime Intellect says long-context agent serving depends on retaining history, scheduling new work, and moving cached state efficiently. It optimized three paths separately: prefill topology and scheduling, compressed KV with a fused attention kernel, and a transfer-friendly cache layout.

AIMuse has released Muse Gadgets, an open-source ESP32 firmware and Linux SDK for building hardware devices that work with Muse. Developers can obtain an API token from gadgets.muse.ai and use a coding agent with the GitHub repo to build peripherals. Guillermo Rauch praised the team's rapid shipping.

AIBaseten tested the MetaInfer skills-only approach by having Claude Code build an inference engine, VibeQwen, for Qwen-3.6-35B-A3B in NVFP4 on a single B200. On single-stream text, VibeQwen decoded 90% faster than a tuned vLLM 0.25.1 deployment (1,792 vs. 943 TPS) and cut time to first token from 28 ms to 12 ms, with a 71% throughput gain at concurrency 32. The author notes this was an outcome-focused run that allowed some numerically different outputs as long as accuracy stayed at or above the BF16 baseline.

Why it matters: The post tests a skills-only inference engine method on a real model and states the speed and accuracy constraints used, helping readers judge how far such automated optimization can be trusted.

AIPerplexity's Computer built a 3D map of nearly 26,000 restaurants and cafes across New York City's five boroughs. Users can search by dish or neighborhood and step inside places such as Peter Luger and Grand Central Oyster Bar. The post frames such projects as ones an agent can run for hours to produce something substantial.

AIMuse has released Muse Gadgets, open-source ESP32 firmware and a Linux SDK for building hardware devices that work with Muse. Developers can get an API token at gadgets.muse.ai and use a coding agent with the GitHub repository to build custom peripherals.

AIAnthropic's ClaudeDevs describes a "You should know" mod that spins off a sideagent to observe Claude's output. The mechanism is detailed in a newer blog post on getting started with Claude Code mods.

AIAnthropic released Claude Code v2.1.288, adding $.ui.selection() for mods, a built-in gh api for cloud sessions without the GitHub CLI, and --max-findings for /code-review. The release also fixes many issues, including mid-response API timeouts, resume and compaction bugs, and auto mode denials and model switching on Bedrock and Mantle.

AIEpoch AI estimates that compute built from projected 2025 to 2027 high-bandwidth memory shipments could support tens to hundreds of millions of frontier AI agents, or billions of cheaper ones. Running nonstop, the top-tier agents would match the working hours of 140 million to 700 million full-time employees, and the central DeepSeek V4 Pro estimate of about 1.9 billion agents would match 8 billion workers.

Why it matters: The estimate converts memory shipments into agent capacity and revenue ranges, showing how hardware supply could translate into labor and sales if demand keeps up.

AIGoogle Research's Cogentic is a multi-agent harness running on Gemini that searches for proofs of open theoretical computer science problems without expert hints. It runs rounds where an orchestrator launches provers, two adversarial verifiers must both accept each draft, and shared disk documents store attempts and verified lemmas. The system produced new results on five open problems in online learning, auction theory, and mechanism design, each checked by domain experts.

AITypeSafe AI shows how to build long-document search by combining its Jev with PageIndex, with no vector database or embeddings required. The project is open source on GitHub under VectifyAI's jev-doc-search repository.

AICursor introduced Rollouts, a tool that writes a monitoring plan and watches changes as they deploy to catch regressions before users see them. When Rollouts detects a regression, it identifies the offending PR and opens an issue, and one click starts a cloud agent to fix it. Rollouts usage credits are included through Oct 3.

AIStratechery's weekly roundup argues that agents could aggregate apps and become tech's most valuable products. It criticizes Meta's Meta Enterprise Platform as a distraction from consumer opportunities, and says OpenAI's Dev Day brought confusing new pricing tiers and overlapping products.

AIGitHub has made the Project HydraFusion research preview available in the GitHub Copilot app and @code, where it orchestrates multiple models rather than acting as a single model. New models from Anthropic (Fable 5.1 and Opus 5.5) and OpenAI (GPT-6.1 Sol) are also now selectable in the Copilot model picker.

AITogether AI's CPO and product team outlined how they use AI to make the whole team more productive, not just individuals. The approach includes a shared context repo readable by any AI harness, cutting half a day of research to about 5 minutes, and evals that test their product the way agents actually use it.

AIStanford HAI faculty affiliate Marco Pavone explains how AI helped close the final 10 percent of the gap to driverless cars, which experts in 2018 said remained. The remaining challenges included handling fog and rain, inconsistent road markings, and safe decision-making. The explanation appears in a Stanford Report article linked in the post.

AIHarrison Chase says LangChain is improving memory in managed Deepagents, noting that memory is difficult to get working well in company settings. The linked background post describes user memory in Managed Deep Agents 0.8, which lets an agent remember the people it works with.

AIHost Vicki Reyzelman of Akamai argues that AI agents can now probe networks, coordinate with other agents, and make purchases faster than organizations can respond. She cites an OpenAI agent that reportedly bypassed security controls while researching Australia's Medicare system, with OpenAI taking 54 days to identify the incident and another month to notify the government. Major model releases are arriving roughly every 17 days, and Meta says its Muse ecosystem has about 1,500 developer connectors.

AIswyx urges the AI community to prioritize security as rogue agents, breaches, and AI-driven attacks multiply. He announces the 2nd AI Security Summit, with Snyk as founding partner and all 2025 partners returning.

AIAi2 released AstaBrief 8B, an open-weights model that turns a research question and retrieved literature excerpts into a cited report, along with its training data. The model runs as Fast mode in Asta, averaging 51.1 seconds per report versus 178.5 seconds for Thinking mode, about 3.5x faster. The post also describes filtering synthetic training data by citation density and building DPO pairs judged by two models that agreed.

Why it matters: The post explains how supervised fine-tuning, preference data, and citation-density filtering were used to build a cited-report model, which is useful for teams training their own models.

AIJoshua Keating argues that keeping a human in the loop on lethal AI decisions is not enough if the humans rely too heavily on AI outputs. He cites a CNN-reported case in which an analyst's AI-assisted report falsely identified a Chinese ship's cargo as nuclear components, nearly prompting a boarding during the Iran war. The piece links this to automation bias and to military AI cases in Gaza and Minab, and warns that AI integration early in a nuclear decision chain is harder to regulate than autonomous launch.