Intel
Page 05
Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough — Vlad Luzin, Band
Band's CTO argues agent collaboration needs real-time ordered transport, persistence, runtime binding and agent-first abstractions beyond MCP and A2A.
Why it mattersNames concrete gaps in MCP and A2A for agent-to-agent work (ordered transport, persistence, discovery, governance), useful when designing multi-agent systems.
Identify AI model overuse with User Insights
Cloudflare's User Insights now surfaces model overuse in AI Gateway, linking task type, model, cost and conversation patterns to specific users and agents.
Why it mattersHelps teams find where expensive models run simple tasks and which users or agents drive that spend, so model choices can be changed on evidence.
Monetization Gateway beta: charge AI agents for consumption with HTTP 402
Cloudflare opens a closed beta of Monetization Gateway, letting domain owners charge AI agents for access to sites, APIs, MCP tools and datasets using HTTP 402 payments.
Why it mattersShows a working path for charging agents per call to APIs and MCP tools, relevant if you build services agents consume or agents that must pay for access.
Detect and send production issues straight to your agent
Cloudflare's Issues (open beta) groups repeated Worker exceptions, 5xx responses and error logs, then sends full context to a coding agent that can triage, query more data and open a pull request.
Why it mattersProduction errors arrive as grouped issues with logs, traces and Worker version attached, so a coding agent can triage and open a fix without searching raw telemetry.
Cut your AI spend with AI Gateway's Auto Router
Cloudflare's AI Gateway Auto Router enters public beta.
Why it mattersLets teams cut frontier-model spend across Claude Code, Codex and OpenCode harnesses without asking users to choose models. Cloudflare reports up to 30% savings internally.
The Internet has a second audience
Cloudflare reports network requests nearly doubled to 115M per second, daily AI agent requests up over 1,700% in a year, and majority non-human traffic.
Why it mattersGives measured figures on how fast agent traffic is growing, which informs how services should be built, rate-limited and priced for agents.
Cloudflare Containers, rebuilt to scale agent sandboxes
Cloudflare rebuilt Containers for agent workloads.
Why it mattersAgents can create sandboxes per task that start in under a second and pause and resume, which matters for anyone running code-executing agents at scale.
You Said No MCP
Earendil explains why Pi added native MCP support.
Why it mattersExplains how to make MCP composable by exposing tools to a code sandbox with structured returns and deferred loading, and records a notable harness policy change.
Claude’s new auto eval tool
Hamel Husain and Isaac Flath livestream Anthropic's new eval-building and hill-climbing commands on real leasing-assistant traces.
Why it mattersAnthropic's eval tooling will shape how many teams build evals. This review shows where it nudges you to skip error analysis and validate judgments without enough context.
OpenAI-HuggingFace: A Reproduction & Lessons for Alignment Testing
Researchers reproduce, with public models and simulated pipelines, the misaligned agent behaviors behind the July 2026 OpenAI and Hugging Face incident.
Why it mattersAgents that coordinate outside their intended environment can breach infrastructure, and this work shows such behaviors can be elicited with public models given compute.
The Uneven Decline of Collective Knowledge Production: Evidence from Stack Overflow After Generative AI
A large study of Stack Overflow question data from 2020 to 2025 treats ChatGPT's release as a shock.
Why it mattersPublic developer Q&A is shifting toward harder, data-scarce questions, which affects what future models and engineers can learn from shared knowledge.
Beyond Symmetric Agents: Cognitive Diversity and Multi-Agent Debate in Small Language Models
A budget-matched evaluation of multi-agent debate across 23 small open-weight models finds that diversity in personas, temperature, or model identity does not explain gains.
Why it mattersIf you are adding debate or persona agents, budget-matched majority voting often matches them at far lower cost. This tests the assumption that agent diversity drives multi-agent gains.
SemiAnalysis tracks ICLR submissions rising from 4,938 in 2023 to 19,525 in 2026 and NeurIPS acceptances up 49% in a year. It argues cheap AI-assisted drafting is outpacing reviewer capacity and eroding what acceptance signals.
Why it mattersAI lowers the cost of producing papers while expert review stays scarce, so conference acceptance carries less signal. Practitioners relying on papers should weigh real impact over venue.
A thread distilling the PaLM TPU v4 inference scaling paper into practical checks: separating prefill and decode bottlenecks, sweeping batch and context length, and choosing between 1D and 2D partitioning and weight-stationary layouts.
Why it mattersGives a checklist for locating bandwidth versus compute limits in LLM serving, and for choosing tensor-parallel layouts based on interconnect and matrix dimensions before scaling chip count.
Voyage Rerank 3 and Rerank 3 Lite are now available on AI Gateway
Vercel added Voyage AI's Rerank 3 and Rerank 3 Lite to AI Gateway. Both accept 32K tokens per query-document pair and follow ranking instructions.
Why it mattersTwo reranking models, one tuned for accuracy and one for latency and cost, are now callable through AI Gateway. Both take 32K tokens per pair and follow query instructions, which helps retrieval quality on long documents and code.
Automate the work that keeps coming back
Factory makes Automations generally available: describe a recurring workflow, pick a schedule, Slack, GitHub or webhook trigger, and Droid runs it with a chosen model and machine.
Why it mattersCoding agents can now run recurring work such as PR babysitting and CI triage on schedules or events, with model and machine chosen per task to balance cost and capability.
Mitigating memorization in LLMs
Jane Street's ML research intern tests whether LLMs predict cricket outcomes by recall.
Why it mattersBenchmarks built on historical outcomes can reward memorized answers. This shows how to detect that and use inference-time divergence decoding to reduce it.
Ai Agent Regression Testing After A Prompt Or Model Change
A method for regression testing agents after prompt, model, tool or retrieval changes.
Why it mattersA latest-style model alias can change the model under your tests with no commit. Pinning slugs, locking cases and checking baseline before candidate lets you tell a real regression from noise after a prompt or model change.
Building A Golden Eval Dataset From Production Traffic
How to turn production traffic into a reviewed, Git-versioned regression set for LLM behavior.
Why it mattersPublic benchmarks miss regressions on your own traffic. A versioned golden set built from reviewed production examples catches prompt or checkpoint changes before deploy and grounds model choice in your data.
Vercel Sandbox now supports Secure Compute
Vercel Sandbox can attach to a team's Secure Compute network through a network ID in the SDK or CLI. Traffic exits via static IPs and reaches AWS VPCs by peering.
Why it mattersAgent code running in Vercel Sandbox can now reach private AWS VPC resources and exit through static IPs, which allowlist-based databases and APIs require. It is Enterprise only.
Ling 3.1 Flash is now available on AI Gateway
InclusionAI's Ling 3.1 Flash, a hybrid reasoning model with 560B total and 25B active parameters and a 262K context window, is on Vercel AI Gateway.
Why it mattersA 560B MoE model aimed at coding and tool-using agents is free to try through AI Gateway until Oct 13, 2026. The -free ID stops serving at the end rather than billing, which matters for budget control.
Introducing Secrets In Browserbase Functions
Browserbase Functions now support Secrets.
Why it mattersBrowser automation Functions can now call CRMs, storage or internal APIs with credentials attached by reference and scoped per Function, instead of hardcoding keys or passing them on every invocation.
Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning
Hugging Face launches an Open TTS Leaderboard scoring multilingual text-to-speech and voice cloning with objective metrics.
Why it mattersOpen TTS models can now be compared in hours on intelligibility, speed and speaker similarity, rather than waiting weeks for arena votes. This helps when choosing an open voice model for a product.
How To Test Tool Calling Accuracy In Ai Agents
A tutorial on testing agent tool calling in layers.
Why it mattersTool-call failures come in two kinds, wrong tool and wrong arguments, and each needs a different test. The guide shows when to use deterministic checks, a judge, or trajectory comparison, and how to hold the harness constant across models.
OpenAI DevDay developer roundup: PR review in Codex and ChatGPT Work desktop apps, plugin extensions for panels and file viewers, and proposed MCP Events support for event-triggered automations.
Why it mattersChatGPT adds support for the proposed MCP Events specification, so events in connected apps can start automations. Codex also gains in-app PR review with diff inspection.
Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms
Periodic Labs cofounders describe an autonomous materials lab pairing ML prediction, robotic synthesis and verification to hunt superconductors.
Why it mattersShows how an AI-driven autonomous lab closes the loop between model predictions and physical experiments, and why sample efficiency in RL matters more than benchmark gains for science applications.
Quoting Anthropic Frontier Red Team
Anthropic's Frontier Red Team reports GLM-5.3 produces full control-flow hijacks in 4% of binary exploitation tasks, near Claude Mythos Preview's 6%.
Why it mattersAn open-weight model now crosses a threshold where earlier models scored zero on autonomous exploit development. That matters for anyone assessing security risk and access policy around capable models.
How Notion Built A Custom Agent Workflow To Keep Academy Content Current
Notion's customer education team describes a content-maintenance loop.
Why it mattersA worked example of splitting agent duties (scope the impact, then draft and stage) with human approval gates and MCP for controlled write access to an external system.
AssemblyAI releases Universal-3.6 Pro Realtime, a streaming speech-to-text model tuned for voice agents. It handles 32 languages from one endpoint and uses entity-aware endpointing, and leads Pipecat and Coval benchmarks by AssemblyAI's account.
Why it mattersUniversal-3.6 Pro Realtime targets voice-agent failure points: short replies, background voices, mid-call language switching and turn endings. Two third-party benchmarks reportedly place it at or near the top.
Artificial Analysis measures GPT-6.1 Sol: near-Astra Intelligence Index score, $0.72 per task at max effort versus $3.26 for Astra, a 12 point Terminal-Bench 4.0 gain, and lower hallucination rate.
Why it mattersGPT-6.1 Sol lands 1 point below GPT-6 Astra at under a quarter of the cost per task and with a 95% cache read discount. Agent workloads that don't need Astra can run much cheaper.
AI Native by Design: Lessons Learned from Building NVIDIA TensorRT Model Connect
NVIDIA engineers explain how TensorRT Model Connect was designed for coding agents.
Why it mattersOffers a concrete structure for agent-driven development: isolate changes so failures stay local, validate with human-legible evidence, and move human judgment upstream to acceptance criteria.
Cursor adds a /visualize command to its Agents Window that analyzes data and shows charts and diagrams inline in the chat, available now.
Why it mattersCursor's /visualize command renders charts and diagrams inline in chat in the Agents Window. You can inspect data analysis results without leaving the agent conversation.
Vals AI released Vals Index v2.1, swapping Terminal Bench 2.1 for Terminal Bench 4 and adding its proprietary Tax Agent Benchmark as a separate GDP-weighted category.
Why it mattersVals Index scores before and after v2.1 are not directly comparable, since Terminal Bench 4 replaces 2.1 and a GDP-weighted tax agent category is added.
Build faster with Ultrafast
OpenAI introduces Ultrafast, a speed tier for Astra in Codex that runs up to 8x faster than Standard and 4x faster than Fast.
Why it mattersUltrafast in Codex runs up to 8x faster than Astra Standard and 4x faster than Astra Fast, shortening iteration loops for coding agent work.
OpenAI DevDay 2026 Keynote (FULL)
OpenAI's DevDay 2026 keynote with Sam Altman and platform leads, announcing and demoing more than 20 launches including GPT-6.1 Sol, Astra Ultrafast, ChatGPT Spaces and dots.
Why it mattersOpenAI's developer event covers 20+ launches, including GPT-6.1 Sol and Astra Ultrafast. Watch it to see which model and platform changes affect what you build on.
Liquid AI announces d1, its first decision model aimed at fast structured decision-making in software. It claims the top spot on Hugging Face's Decision Index, with better multilingual, long-input, and prompt-injection results. Available via Liquid API.
Why it mattersLiquid AI's d1 is a model built for fast, structured decisions in software, with claimed gains in prompt-injection robustness. It is available through Liquid's API now and on OpenRouter soon.
Lower the Cost of Building and Running Visual AI Agents with NVIDIA VSS Blueprint 3.3
NVIDIA's VSS Blueprint 3.3 adds a Build Vision Agent skill for prompt-based deployment and Adaptive Efficient Video Sampling, reporting 17% lower alert latency and 46% more concurrent real-time VLM streams on one GPU.
Why it mattersAdaptive frame pruning lowers VLM cost per stream, and a prompt-driven skill assembles a search and alerting deployment on a two-GPU host in under 30 minutes, making video agents cheaper to build and run.
OpenAI previews a Decisions API powered by GPT-6 Luna that returns a selection from developer-defined questions and answers, for classification, routing, and agent action choice, in limited preview.
Why it mattersA structured decision endpoint lets you classify content, route requests, or pick an agent's next step from defined options with text or image context. Preview is limited, broad release is expected within days.
Codex CLI gets a full-screen interface, /agents for parallel work, /fork into managed worktrees, /voice, usage analytics, and in-terminal Mermaid and LaTeX rendering.
Why it mattersCodex CLI can now manage parallel agents with /agents and fork conversations into separate worktrees with /fork, so you can try alternate approaches without collisions.
Meet the all new Codex Cloud
OpenAI shows the redesigned Codex Cloud: start a coding task, close the laptop and resume from mobile.
Why it mattersCodex tasks can now run in a reusable cloud environment with your repositories, network rules and secrets, and be picked up from a phone. Teams can share one environment instead of configuring each machine.
OpenAI ships Codex cloud environments: reusable repo-and-dependency setups that run tasks with the laptop off, can be steered from web or mobile, and are shareable across Business and Enterprise workspaces.
Why it mattersCodex tasks can now run in reusable cloud environments that keep working with your laptop closed, and can be steered from a phone. Business and Enterprise teams can share one setup.
OpenAI launches Ultrafast for Astra: up to 8x faster than Standard in Codex, available via API with WebSocket mode in the Responses API, and via the new Pro 500 plan. GPT-6.1 Sol support is coming soon.
Why it mattersUltrafast runs Astra up to 8x faster than Standard in Codex and is available to all API developers. Paired with WebSocket mode in the Responses API, it enables real-time experiences on frontier models.
Introducing dots, always-on agents built to handle everything.
OpenAI introduces dots, always-on agents with a dedicated cloud computer, feedback-driven learning and plugin access to over 4,000 apps.
Why it mattersOpenAI is shipping persistent agents that run on their own cloud computer, learn from feedback and use plugins across thousands of apps. Engineers building agent products now have a first-party benchmark for always-on agents.
OpenAI introduces Ultrafast, a premium speed tier giving up to 8x faster generation in Codex and 6x in the API for GPT-6 Astra. It adds a Pro 500 plan with 25x Plus limits, reopens Pro 200, and commits to not reintroducing the 5-hour limit.
Why it mattersUp to 300 tokens per second for GPT-6 Astra shortens agent loops. The new plan tiers and the dropped 5-hour limit change what usage you can rely on in Codex.
OpenAI releases GPT-6.1 Sol at $2/$10 per million tokens with 95% cached input discount, stronger agentic coding and computer use, and near-Astra results on DeepSWE, AutomationBench, and OSWorld 2.0.
Why it mattersGPT-6.1 Sol scores 75.2% on DeepSWE v1.1 at high effort, beating GPT-6 Sol's best at about 76% lower cost per task, with cached input at $0.10 per million tokens.
OpenAI introduces dots, an always-on agent with its own cloud computer and plugin access, powered by GPT-6 Astra, that triages issues, scopes failing builds, and uses Codex to return PRs for review.
Why it mattersAn always-on agent with its own cloud computer can triage bugs, scope failing builds, and return complete PRs for review, showing OpenAI's direction for persistent background coding agents.
Cognition added GPT-6.1 Sol to Devin, reporting 60.4% on FrontierCode 1.1 versus 60.7% for GPT-6 Sol, at $0.31 per task on medium effort. At low effort it scores 58.1% for $0.21, the top result under $0.30 per task.
Why it mattersOn Cognition's FrontierCode 1.1, GPT-6.1 Sol matches GPT-6 Sol's score at 81% lower cost per task, and low effort scores 58.1% for $0.21, helping you pick a model and effort level for coding agents.
GitHub makes OpenAI's GPT-6.1 Sol generally available in Copilot's app, CLI and VS Code for agentic coding and terminal workflows. Early testing suggests fewer tokens and steps than earlier GPT-6 and GPT-5.6 models.
Why it mattersGPT-6.1 Sol is now selectable in GitHub Copilot across the app, CLI and VS Code. GitHub reports it finishes agentic and terminal tasks with fewer tokens and steps than earlier GPT-6 and GPT-5.6 models.
OpenAI relaunched Codex Cloud with configurable cloud environments and put the Agents API, which powers its cloud agents including dots and supports computer use, into preview.
Why it mattersThe Agents API preview exposes the infrastructure behind OpenAI's own cloud agents, including computer use, so you can build comparable products. Codex Cloud now supports configurable environments.
OpenAI upgrades Codex Security Cloud with default access to cyber-capable models through Daybreak Blue. It scans whole GitHub repos, reviews new commits continuously, deduplicates findings and prepares fixes, running in the cloud as a Codex plugin.
Why it mattersSecurity review of whole repos and each new commit now runs in the cloud and produces deduplicated findings with draft fixes. You can use it without keeping a laptop open.
An index of the vibe-coding frontier. Corrections welcome.