Agent Infrastructure

Spline says Hana V2 lets agents generate through MCP or its in-app tools. The release adds PBR GLB and GLTF import, SVG and PDF export, a WebGPU canvas, and improved vector editing.

A Tembo user demonstrated real-time frontend iteration in cloud Linux VMs with databases in the same environment. Apps can be tunneled from a VM and shared by link, while nested virtualization is being added.

JFrog released the free Boost CLI, which compresses shell output before it reaches coding agents including Claude Code and Codex. Its Terminal-Bench 2.0 report claims a 13.5% cost reduction across 89 tasks.

A game developer documents a Grok Bot workflow that creates consistent game art asset sheets with Grok Imagine, evaluates results, and exports transparent PNGs. Separate bots summarize Slack activity and send bug reports to a coding agent for reviewable fixes.

OpenClaw says it redesigned its web interface and native apps around visualizations for goals, compaction events, and team identities. The project says the web UI is nearing a stable release and supports local hardware and Ollama.

Spline rebuilt its web-based 3D editor with Agent Mode and an MCP connection for its desktop app. V2 also adds a WebGPU engine, PBR and HDR workflows, and scripting for code-enabled scenes.

Users report deploying Grok Bot for churn recovery, support inboxes, refunds, expense reduction, and sales prospecting. One report says it found $175 a month in subscriptions to cancel after obtaining approval.

Hailuo says MiniMax Design uses agents to plan and execute video production around MiniMax H3. Creator examples show agents deriving camera movement and pacing from a reference image before writing an H3 prompt.

Intangible connects to Claude Code, Claude Desktop, or ChatGPT through MCP so an agent can build scenes, compose cameras, and preview renders. The beta leaves creators to finish directing inside the scene.

Cursor’s Origin hosts repositories, reviews pull requests, runs agents, and deploys through Vercel. GitHub repositories can sync while GitHub remains the source of truth.

Tutorials and user reports show Grok Bot handling web research, inbox cleanup, travel planning, and game installation through a persistent cloud computer. The trade-off is a focused, responsive experience against lower flexibility, a reported $200 monthly starting price, and lag during gameplay.

Stages says V5 gives VIDX and SIGNAL shared editor intelligence, durable command history, deterministic CUE actions, and safe replay. Stages Connect can control TouchDesigner, Blender, After Effects, and Unity through MCP.

Benjamin Cherny said Anthropic trains Claude against prompt injection and that Claude Code’s model, probe, and auto-mode layers can reach 0% attack success in next week’s defaults. The claim comes from an Anthropic engineer rather than an independent benchmark.

Boris Cherny said Anthropic has landed large P99 RSS improvements for Claude Code while still collecting CPU and memory reports across CLI and Desktop. The issue remains an active performance investigation requiring machine, OS, version, and task details rather than a fully verified fix.

Anthropic’s Boris Cherny said Claude Code users are not banned for using other model harnesses through proxies. He pointed to a classifier issue, unblocked one escalated account, and said false positives are being reduced.

Anthropic says Claude Code Pro, Max, and Team users will default to Auto Mode on Aug. 14. Its tool-call classifier reportedly caught 89% of dangerous commands, versus 14% for manual approval, after prompt-injection testing.

Developers proposed agent-specific browser auth for Codex, Devin, Claude, and Cursor so agents can use selected passkeys, TOTPs, and passwords without exposing full vaults. Related posts warned that Claude Connectors may broaden tool access.

Developers shared workflows that turn idle AI accounts into scheduled assistants for commits, research, repo cleanup, and notes. Shared designs add spend caps, Apple Notes routing, Obsidian output, and coworker controls.

Allie Miller described using Codex Voice Mode for trackers, email, reminders, scheduling, and document work. Min Choi’s ElevenLabs setup shows a browser agent that listens, streams replies, and handles interruptions.

Shared workflows used Codex, Claude Code, and Fable for release testing, inbox triage, product-plan review, and adversarial code review. OpenClaw’s QA prompt split testing across 12 subagents.

Anthropic’s developer update adds per-agent effort settings, seeded session creation, webhooks, sub-agent event streaming, and up to 500 skills per session. The release gives teams finer controls for managed agent runs.

Google says Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber target faster, cheaper agent workloads. Early tests cite lower token use and stronger coding results.

Peter Yang shared Thariq’s explanation that Anthropic cut Claude Code’s system prompt by 80% as newer models need fewer examples and more room to reason. The same discussion covered /loop, /goal, workflows, and long-running agents.

ClaudeDevs said Pro, Max, Team, and seat-based Enterprise users keep 50% higher Claude Code weekly limits through August 19. Users had warned the promo’s end would sharply cut effective usage.

LLMJunky said Codex used Palmier with FFmpeg to choose frame ranges, speed up a clip, add overlays, animate an icon, and place text. He still handled some zoom effects manually.

Levelsio used Claude Code on a VPS to SSH into a cloud Mac Mini, build an iOS app headlessly, and stream serve-sim into a browser. Follow-up posts documented the tunnel and serve-sim setup, making the workflow reproducible for remote iOS testing.

Convos launched an agentic messenger where tagging posts, screenshots, or ideas builds Hermes agents for group chats. Demos show Fable-in-chat, weekly briefings, payment-chasing agents, travel planning, and a no-signup builder harness.

Official demos show Claude via Higgsfield MCP building small apps such as Ad Studio, NutScan, a handwriting-to-font tool, and a hand-tracked solar system app in one session. The demos frame Higgsfield Apps as a one-prompt path from idea to deployable mini app.

Kitze described using a $60 Hetzner server, four $200 Codex accounts behind codex-lb, self-hosted Paperclip workspaces, and a GPT-5.5 manager to run about 30 isolated tasks. The setup routes parallel coding work across isolated workspaces and paid Codex accounts.

Figma added MCP connectors so the Figma agent can reach apps like GitHub, Slack, Notion, Atlassian, Granola, Hex, and Dovetail. Testers can start trying connected code and video workflows as the beta rolls out.

Stages AI posts described a server-side runtime that keeps working off-tab, then added CUE continuity, Signal editor previews, drop zones, and provenance inside CASTING. The update matters because Stages is positioning itself as an orchestration layer for characters, prompts, and postproduction rather than a simple chat interface.

Practitioner posts describe loop-based agent systems for coding, PR, sales proposals, and app building, including Kun Chen’s 40-PR-a-day setup, a nine-part vertical-agent framework, and Netlify agent runner builds. Builders can use these patterns to move from single prompts to orchestrated systems with planning, memory, evals, and human checkpoints.

Cursor launched public profiles with handle claiming and a team setting to turn profiles on. The new pages give AI coders a shareable public identity layer for agents and activity inside Cursor.

Anthropic released the ant CLI so Claude Platform APIs, file uploads, and Managed Agents sessions can run from the terminal, then updated Claude Code so /fork starts a background agent with the same context and prompt cache. Teams can use it to script agent runs, inspect traces, and hand work between Claude Code and the platform.

OpenClaw maintainers said Microsoft showed the OpenClaw gateway at Build and tied it to Windows Execution Containers for native sandboxing. The observability and verifiable-workspace features push the project closer to enterprise computer-use deployments on Windows.

Codex users shared 56-hour task runs, PM-to-PR workflows, and a new black-box session recorder for tracking drift, token use, and incomplete responses. The longer autonomous sessions matter because browser auth gaps, passkey failures, and tool-selection bugs become real blockers once Codex is used beyond quick code generation.

Google said AI Studio can now build Android apps directly, add managed agents, and export projects into Antigravity with one click. Paired with Antigravity’s new 2.0 app, CLI, and SDK, the stack moves Google’s prompt-to-product workflow beyond chat demos into runnable apps and multi-agent builds.

Skilled launched a CLI and TUI that scans installed skills across Claude Code, Codex, Droid, OpenCode and Grok Build. It surfaces dead skills, single-project dependencies and usage by agent or project, so teams can clean up skill sprawl.

Anthropic said paid Claude plans will get a dedicated monthly budget for Agent SDK, claude -p, GitHub Actions, and Agent SDK apps starting June 15. Keep chat use and programmatic use separate, and note the temporary 50% weekly Claude Code increase through July 13.

Holaboss launched holaOS Beta 0.1 with permanent workspaces, sub-agents, and a dashboard for recurring tasks and review loops. Use it if you want persistent project memory instead of reset-every-run agent chats, though the evidence is mostly launch-thread documentation.

Anthropic opened Agent View as a research preview, giving Claude Code one control pane for parallel sessions, skills dispatch, and quick replies. The change makes multi-session supervision a native workflow instead of a terminal-tab workaround.

Anthropic introduced dreaming as a research preview in Claude Managed Agents alongside multiagent orchestration, rubric-based self-improvement, and webhook updates. Sub-agents now share a container and filesystem, so teams can coordinate longer-running work and manage memory across sessions.

Anthropic said a SpaceX partnership will add compute capacity, and it doubled Claude Code 5-hour limits for paid plans. It also removed peak-hour reductions for Pro and Max and raised Opus API limits; the change should reduce throttling for heavy users.

SubQ launched a sub-quadratic sparse-attention model with a 12 million token context window and opened early access alongside SubQ Code. The company claims 52x faster 1M-token performance than FlashAttention and under 5% of Opus cost, putting long-context coding workflows into a new price and latency band.

Weekend builder posts showed OpenAI Codex using /goal to keep working across turns, with Linux clients and ephemeral runner tools extending longer sessions. It matters for vibe-coders packaging Codex into unattended loops, but usage limits and community wrappers still vary by plan and platform.

Cross-author demos showed Hermes using self-rewriting skill files, timeboxed subagents, and recurring brief workflows that improved over repeated runs. It matters because creators and vibe-coders can compound agent behavior across sessions, though the evidence still comes from user-run setups rather than a full official product brief.

Multiple practitioners showed Codex reviewing every main-branch commit, spawning fix loops, and opening browser sessions when APIs or web apps blocked the normal path. The workflow matters because Codex is being used as a browser-native coworker for coding, writing, analytics, and media plugins, but the pattern is emerging from user experiments rather than a formal OpenAI release.

OpenClaw contributors posted a voice-persona feature and fresh performance numbers that cut first output from 1s to 43ms. Separate posts describe 300-user sandboxed deployments and stronger PR, CI, and testing workflows, pointing to team-scale use beyond hobby demos.

Anthropic published a post-mortem on Claude Code regressions, said the problems lived in the harness rather than the models or API, and reset subscriber usage limits after fixes. The update matters for long agent sessions because recent complaints centered on reliability, wasted effort, and broken trust.

Elon Musk said X API access is now available through OpenClaw, while users posted travel-assistant and always-on marketing setups built around the tool. The new access broadens what OpenClaw agents can automate, but most concrete examples still come from operator threads rather than product docs.