The Archive

Robot agents infer unknown hardware constraints of partners via observation, then coordinate zero-shot on new manipulation tasks.

AutoCompact trains coding agents to autonomously decide when and what context to compress during long-horizon repository-level tasks.

LLM agents develop persistent source-specific competence through repeated access to authoritative knowledge sources on knowledge-intensive task sequences.

Real-world enterprise data science and analytics workflows require reasoning across dozens of tables, performing statistical analyses, and acting on the results. Established text-to-SQL benchmarks evaluate query generation alone, and audits have found their answer keys frequently wrong. Because real enterprise warehouses are too sensitive to release, these benchmarks are built on public datasets where a business event fits in a single table. We introduce Argo-Bench, an evaluation framework comprising 210 data science and analytics tasks. Drawing on public data, peer-reviewed industry literatu...

Mimir physics-grounded LLM agent learns long-horizon irrigation control with compounding errors and no reset.

Brian Chesky on making Airbnb agent-friendly, the state of consumer AI, and why the world needs an AI-native operating system.

A pop-up shop for "dots," a personal assistant agent at OpenAI DevDay 2026. (Photo by Heather Diehl/Getty Images) | Getty Images At OpenAI's annual DevDay conference, the company pulled out all the stops to compete with its rivals - primarily Meta, whose Muse AI agent platform has seen early runaway success. CEO Sam Altman walked onstage to cheers and announced Dots, a "real-deal AI" agent powered by GPT-6 Astra, "inspired by the cool agents that we all watched in movies growing up." Similar to Muse, Dots aims for disarming cuteness, taking the form of colorful, personalizable blobs with eyes...

The startup helps developers build AI agents that work over iMessage, SMS/RCS, email, and other messaging platforms. It's a bet that consumers will increasingly use agents instead of downloading apps.

PACE system enforces safety for tool-using LLM agents by tracking provenance and mediating tool calls before execution.

Claim-specific verification audit for modular agents replaces aggregate scoring with evidence-based component evaluation.

SCOPE-AD framework for cost-aware Alzheimer's disease diagnosis using ordinal-belief planning and energy-based models to guide sequential test acquisition.

DeFA traces failure propagation in LLM agents by constructing dependency graphs to localize multi-step errors separated by execution steps.

Training approach enables language agents to internalize control decisions (revision, stopping) from harness-supported trajectories, reducing runtime dependency.

Matthew Green warns that AI agents in separate sandboxes could enable cross-agent worms via shared package caches, email, and Slack—a novel supply-chain security threat.

Flow Engineering, which is bringing AI agents to hardware design, also landed Roelof Botha as an angel investor and board member.

OpenAI's "Decisions API" is a Jev clone that confirms the importance of fast, cheap intelligence.

Sam Altman onstage at OpenAI’s DevDay 2026. | Image: Hayden Field / The Verge While AI has made plenty of inroads on people's phones and computers, it's largely failed in dedicated devices. But over the next year, two major AI companies, Meta and OpenAI, will attempt to change that. And they're apparently betting on a similar path: testing appetite for physical hardware with their cutesy software agents. "Hardware is hard" is a tech industry mantra. Dedicated AI devices like the Friend and Humane AI Pin have inspired mainly frustration and backlash, which may only worsen as public anger again...

WorldAuditBench benchmark for detecting anomalies in 3D environments using multimodal agents combining vision and navigation.

DynaHarness couples semantic reasoning with physical execution via shared contracts to enable robot agents to revise failed capabilities iteratively.

Study on autonomous ML engineering agents finds primitive LLM coding agents with direct environment access outperform complex multi-agent orchestrators.

PivotOPD method recovers from early pivotal mistakes in multi-turn agents via on-policy distillation; tested on Qwen 8B–235B models.

cua-speedrun benchmark standardizes reproducible speed evaluation for GUI computer-use agents to address cost and latency deployment barriers.

ComputerSD applies online self-distillation with privileged rescoring to CUA training, addressing guidance misalignment and probability shift problems.

PhantomEnvironments trains LLM agents via RL using rule-based synthetic fictional worlds with zero marginal cost, avoiding hallucinations and benchmark contamination.

GGSD discovers interpretable, playable robot motor skills via self-play in games, enabling human control of embodied agents.

Persistent Context Graphs compress LLM agent memory using attention history signals to reduce prefill cost and stay within context windows.

Founded by Apache Flink veterans, the startup is taking on workflow heavyweight Temporal.

Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face, OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised serious questions…

Predictive 4D belief framework enables navigation agents to infer and revise moving target locations in dynamic environments.

DAGent uses directed acyclic graphs with evaluate-then-grow planning to dynamically scaffold deep research tasks for multi-agent systems.