Agent runtime infrastructure

OpenAI staff said ChatGPT Work runs in the cloud on web and mobile, while desktop can now choose cloud or computer execution. The clarification followed confusion about closed-laptop and local-environment behavior.

New releases and practitioner posts described graph-shaped agent runtimes instead of simple loops. Slate, LangChain's open-sourced agent factory, and AI SDK's Python package each add pieces for durable orchestration.

A Stanford-linked Shepherd thread described rollback for long agent runs that restores files, packages, dev servers, and process state beyond git-style rewinds. Replies flagged KV-cache warmth and registered database inverse steps as cost and recovery boundaries.

AI SDK added HarnessAgent as a common interface for Pi, Claude, Codex, OpenCode, and other harnesses. Use it to run local or cloud software-factory jobs through official SDKs while subscriptions cover token usage.

Vercel shipped FUSE-based Sandbox mounts for S3 and network filesystems and opened Agent Runs through MCP and CLI. Use it to connect remote state, sandbox execution, and agent-readable Eve traces for self-improving workflows.

Mastra published an OpenUI guide, CopilotKit showed Fable 5 agents rendering AG-UI React components, and OpenUI Adam targeted Vercel Eve. The cluster creates a shared component-streaming surface across React, design systems, and filesystem agents.

harbor exec launched an agentic-map-reduce CLI installed with npx skills add harbor-exec. Use it to run sandboxed agents for trace analysis, session mining, search, and rollout aggregation.

Browser Use CLI 3.0 shipped direct Chrome DevTools Protocol control through browser-harness with a 6× smaller context path. Try it with Claude Code, Codex, cloud browsers, or local Chrome sessions to cut browser-agent context overhead.

Apify added more than 20,000 Actors to the x402 flow, letting agents pay in USDC and run tools on demand through HTTP 402 responses. That gives agents a way to buy web automation tasks without pre-provisioned API keys or a manual checkout step, so builders can test paid tool use directly.

Vercel extended the AI SDK Harness API to cover OpenCode and Deep Agents, adding more agent runtimes to the unified interface introduced in AI SDK 7. The change matters because apps can swap supported runtimes without rewriting integration code, though ACP is still awkward for some cloud deployments.

Rivet released agentOS v0.2.0, a Rust rewrite of its WebAssembly-based sandbox and orchestration stack with multiplayer workflows and one-prompt deployment. The release targets self-hosted and cloud agent runtimes, and Rivet claims 1738x lower cost than SaaS sandboxes.

Vercel shipped AI SDK 7 with approvals, durability, telemetry, and other production agent primitives. Early adapter feedback points to breaking changes and migration work for SDKs that wrap the old APIs.

Latitude released an open-source platform for monitoring AI agents in production, with plain-English trace search, repeated-failure clustering, and MCP access from coding agents. That gives teams a self-hostable way to inspect token burn, surface recurring failures, and turn production traces into evals and fixes.

Google put the Interactions API into GA as the new default for Gemini, adding background execution, managed agents, remote sandboxes, and multimodal tools. Builders now get one stateful interface for models, long-running jobs, and future Gemini Omni support.

Sakana launched Fugu Ultra on AI Gateway and published a technical report, with early testers sharing mixed results. Reports mention polished outputs on some tasks, but also 30-minute runs, uneven coding quality, and much higher cost than GLM-5.2.

ComputeSDK published results from its 2026 100k Scale Invitational after weeks of reruns and infra tuning across Modal, Tensorlake, Northflank, Declaw AI, E2B, and Isorun. It matters because sandbox and agent infra claims now have a shared public concurrency target instead of vendor-specific load demos.

Databricks open-sourced Omnigent, a meta-harness that runs Claude Code, Codex, Cursor, Pi, and custom agents in one live session with a collaborative web UI. The release centralizes supervision, cost control, and cross-agent review instead of splitting work across separate tools.

ENPIRE launched a physical autoresearch setup that gives eight Codex agents robots, GPUs, and real-world APIs for tasks like zip ties and part sorting. It matters because it moves long-horizon agent evaluation from browser-only loops into embodied experimentation with explicit safety controls.

Factory 2.0 expands from coding agents to end-to-end software factories spanning tickets, code, tests, deployments, and incidents. The release makes feedback loops and model routing first-class parts of engineering automation.

Vercel opened a preview that lets Functions run for up to 30 minutes on its Fluid microVM compute platform. Use it for longer-running server tasks without moving to a separate runtime product.

AI SDK canary added HarnessAgent, a unified abstraction that runs Claude Code, Codex, and Pi in sandboxed sessions with AI SDK-compatible streams. One integration can now target multiple agent harnesses without separate model-specific plumbing.

OpenAI said it will acquire Ona and fold its secure cloud execution and orchestration stack into Codex. The change targets agent jobs that need to keep running for hours or days after the original laptop session ends.

Vercel made the skills.sh API generally available, exposing more than 600,000 skills as a registry-style service for agents and platforms. The launch gives teams a discoverable capability layer for reuse across agent surfaces.

TanStack AI added MCP support for single or multiple servers, standalone clients or pooled servers, and a CLI for type generation. The release gives app builders a typed integration path for MCP-managed tools inside chat and agent workflows.

Weaviate introduced Engram, a dedicated agent memory service with async writes, semantic topic grouping, tenant scopes, and composable pipelines. It matters because teams can add a hosted memory layer for agent stacks without stitching custom memory workflows into each application.

Browser Use launched synced cloud profiles for logged-in sessions, added geo-targeted proxies, and showed a 484-browser startup demo that finished in under two seconds. The update matters because hosted browser agents can now keep authenticated state and regional routing without custom session-management work.

LangSmith added sandboxed execution, spend-aware gateway routing, and Engine to surface recurring agent failures from traces. The bundle gives teams one place to run agents, control token spend, and turn production issues into debugging and eval loops.

CopilotKit shipped v1.59.2 with threads, Vue packages, a React Native SDK, and updated AG-UI building blocks for fullstack agent apps. The release makes it easier to ship Cursor- and Claude-like interfaces, with new work extending generative UI into Slack, Teams, and other chat surfaces.

Multiple agent-infra vendors shipped copy-on-write branches, checkpoints, snapshots, forks, or rollback primitives on the same day. That matters because long-running agents can now explore, retry, and recover state without relying only on Git or full sandbox rebuilds.

NVIDIA released Cosmos 3 as an open omnimodel family with 16B and 64B variants, plus code, datasets, and a coalition around physical AI. The release matters because it ships with serving support and top open-weight image and video rankings, so teams can use it beyond a research teaser.

OpenAI made GPT-5.4, GPT-5.5, and Codex generally available through Amazon Bedrock. AWS shops can now use OpenAI models inside existing IAM, compliance, and procurement workflows instead of adopting a separate vendor stack.

Perplexity replaced one-shot search calls with Search as Code, a Python-based search runtime in its Agent API that is also now the default in Computer. The change matters because agents can batch, rank, filter, and aggregate search steps inside code, and Perplexity says the system scored 0.386 on WANDR versus 0.152 for the next system.

Microsoft and NVIDIA unveiled RTX Spark systems, including Surface Laptop Ultra and DGX-class Windows hardware, with 128GB unified memory and 1 PFLOP FP4 local AI. Day-one support from Hermes Agent, vLLM, Ollama, and Unsloth makes the launch useful for local inference and fine-tuning, not just a PC refresh.

OpenAI shipped a Python SDK and app-server support for Codex with thread creation, streamed turns, session resume, image inputs, and sandbox controls. That gives teams a supported way to embed Codex inside internal tools and automation instead of driving it only through the CLI or desktop app.

Browser Use rebuilt its runtime around a custom Chromium fork, Firecracker fork, and custom Linux kernel, claiming $0.02 per hour pricing with subsecond cold starts. The shift targets the infrastructure bottlenecks behind browser agents rather than model quality alone.

Independent developers shipped sidecars that let Claude Code, Cursor, and Codex share memory, hot-swap model providers, package local projects as apps, and automate browser QA. Try these reusable tools if you want memory, routing, QA automation, and app packaging outside editor-specific features.

CopilotKit shipped an AG-UI integration that streams Claude Agent SDK agents into web and mobile frontends with generative UI and approval checkpoints. The adapter lets teams embed terminal-first Claude agents in React, Vue, Angular, and React Native without rewriting transport or state plumbing.

Builders released a chat-first Web UI and a multi-agent Control Room template around Hermes Agent, while core updates cut read_file input tokens by 14% and fixed TUI startup hangs. Use the new controls to manage local multi-agent setups while reducing routine token burn.

Prime Intellect launched Hosted Evaluations to manage harnesses, sandboxes, and rollout inspection for model testing. The service packages eval infrastructure while still supporting local runs against arbitrary engines, so teams can centralize testing without losing flexibility.

Vercel Sandbox can now build and run Docker containers, persist images and installs across sessions, and host databases or full apps inside the sandbox. That broadens what coding agents and preview environments can validate without leaving Vercel.

Gemini Managed Agents can spin up a sandboxed Linux environment with code execution, web access, and file I/O from one API call, and early examples now include W&B and LlamaIndex workflows. That gives builders a higher-level runtime for long tasks while third-party templates start to define the first production use cases.

Nous Research released Hermes Agent v0.15.0 with skill bundles, MCP Catalog, new model support, and major performance and security work. The update cuts load times 50%, speeds session search 750x, and adds Bitwarden plus prompt-injection defenses.

Trajectory launched a platform that turns agent traces and user corrections into post-deployment model updates instead of prompt-only fixes. Baseten and Tinker described live A/B post-training, 397B-model deployment work, and an off-policy recipe for stabilizing the loop.

Hermes Agent added a built-in MCP Catalog while separate builders shipped Qwen3.7 Max support, Venice private-model workflows, and Krea 2 image generation. The cluster shows Hermes moving beyond a single-model assistant toward a broader agent shell with tool, model, and media providers.

Cua Driver said its Windows backend is now stable, letting Claude Code, Codex, Hermes, or custom agents drive real Windows apps through MCP or CLI. The release targets Windows-only line-of-business software while keeping the desktop usable with multi-pointer support.

Firecrawl is now available through Vercel Marketplace and Agent Marketplace for apps and agents that need live web data. The integration reduces setup friction for teams adding scraping, search, and structured retrieval to deployed AI workflows.

A new MeMo paper and several community memory systems converged on keeping knowledge outside the base model through recipe files, semantic and autobiographical stores, and background reconsolidation. The pattern matters because engineers are treating context loss as a systems problem instead of only asking for larger context windows.

Datasette 1.0a30 introduced a slash-triggered Jump To menu plus a hook for plugin-supplied search items. Simon Willison used it in datasette-agent 0.1a4 to start agent chats from the same menu, so plugin authors can wire in their own actions.

LangChain opened a private beta for Managed Deep Agents, a model-agnostic deployment layer built on deepagents with durable execution, sandboxes, and a context hub. The release turns deep-agent rollout into a single config-and-deploy flow and adds an auth proxy boundary for agent actions.

Anthropic added self-hosted sandboxes in public beta and MCP tunnels in research preview to Claude Managed Agents. Use the new options to keep agent execution inside your perimeter or private cloud and reach internal MCP servers without public exposure.