Skip to content

#MCP/工具调用

3 today

Apr 14Tuesday

X · @dotey

Rather than AI First, this is really Software Engineering First

The post argues “AI First” is an engineering problem: if AI writes code in 2 hours, review, testing, deploy, monitoring, and rollback must also run automatically, with humans kept at key decision points. Its concrete prerequisites are automated tests, CI/CD, A/B testing, production monitoring, task management, and a clear architecture; without them, a 25-person team just shifts bottlenecks from coding to QA and ops. The real boundary is use case fit: API services, data platforms, and internal tools fit better than complex UI, core products, or high-security systems.

Why it matters: The author argues AI First's bottleneck shifts from writing code to QA and ops, and lists hard prerequisites like automated test coverage, deployment monitoring and rollback for teams to gauge their own engineering base.

X · @dotey

Vercel open-sources Open Agents, a reference implementation for enterprise coding agent platforms

Vercel open-sourced Open Agents as a forkable reference for enterprise coding-agent platforms, with a three-layer architecture and features like voice input and PR creation. Its key design keeps the agent outside the sandbox and uses tools such as file I/O, shell, and search to control execution; the post also cites Anthropic Managed Agents pricing at $0.08 runtime per hour and $10 per 1,000 web searches. The part to watch is the agent-sandbox split, not the packaging choice.

Why it matters: Vercel open-sources Open Agents, a forkable reference implementation for enterprises building coding agents, with a three-layer split of agent and sandbox that lets the agent remotely drive an isolated execution environment.

最佳拍档 (BestPartners)

Meta-Harness: Can harness engineering code self-iterate? A Stanford paper analysis

Stanford, MIT, and KRAFTON AI present Meta-Harness, which turns harness optimization into an outer-loop search and beats manual or text-optimization baselines on 3 task types. The system uses a coding agent to inspect filesystem history; after 10 search iterations, the data exceeds 10 million tokens, and on online text classification it matched OPRO’s 60-iteration result in 4 iterations while reaching 75.9% average accuracy on 5 OOD datasets. The key point is full-feedback retention rather than compression; the paper also reports about 20 TerminalBench-2 iterations at a total cost of a few hundred dollars.

Why it matters: Stanford, MIT and KRAFTON AI propose Meta-Harness, which turns tuning the code around a large model into a search problem, with a coding agent reading past runs and rewriting the code itself.

X · @dotey

Developer Can Vardar says disabling telemetry in Claude Code cuts prompt cache from 1 hour to 5 minutes

Can Vardar said disabling telemetry in Claude Code drops prompt cache from 1 hour to 5 minutes; Anthropic engineer Boris Cherny said the client then falls back to the 5-minute default because experiment flags stop working. The post says 1-hour cache costs more to write and less to read, so value depends on reuse; Anthropic plans env vars to force 1 hour or 5 minutes.

Why it matters: A developer found disabling telemetry cuts Claude Code's prompt cache from one hour to five minutes, and an Anthropic engineer explains the cost tradeoff.

Apr 13Monday

最佳拍档 (BestPartners)

2027 Is the Enterprise AI Singularity Year: Sundar Pichai on 10 Years as Google CEO, Transformer and Search

Sundar Pichai said in a Stripe interview that Alphabet plans $175B-$185B in 2026 capex and that 2027 will be the breakout year for enterprise AI agent workflows. He said Google cut Search latency by 30% over five years while adding AI features, manages teams with 10 ms or 30 ms latency budgets, and sees 2026-2027 constrained by wafers, memory, power, and permitting. The point to watch is not search replacement but search evolving into an agentic manager, while TPU allocation has become Google's scarcest internal resource.

Why it matters: In an interview, Google CEO Pichai picks 2027 as the breakout point for enterprise AI agent workflows, with AI moving from a coding productivity tool to the intelligence core for non-technical roles.

Apr 12Sunday

X · @Yuchenj_UW

MiniMax M2.7 is open-source!

MiniMax open-sourced M2.7 and said its research agent now handles 30%–50% of the R&D workflow. The post says the agent covers literature review, experiment orchestration, log debugging, code fixes, and merge requests; M2.7 also rewrote its own harness for 100+ automated rounds, with a 30% gain on internal coding evals.

Why it matters: MiniMax open-sources M2.7 and publishes internal experiments: its research assistant agent already handles 30% to 50% of the R&D process, and the model rewrote test scaffolding on its own for over a hundred rounds.

Apr 11Saturday

X · @dotey

Anthropic launches Claude Managed Agents beta; Michael Cohen explains secure third-party key management for agents

Anthropic added Vaults to the Claude Managed Agents beta to manage each end user's third-party credentials with a per-user vault_id and automatic injection at session runtime. The post shows a three-step flow—create a Vault, bind credentials to an MCP server address, and pass vault_id when creating a session—and prices CMA at token usage plus $0.08 per session-hour. The key design is isolation: credentials never enter Claude's context window, code runs in a sandbox, auth goes through a dedicated proxy, and the harness cannot access secrets.

Why it matters: Anthropic's Claude Managed Agents beta adds Vaults, so developers can store third-party credentials per end user and pass only a vault_id in the session.

X · @OpenAI

OpenAI says an Axios third-party library security issue prompted macOS app certificate updates

OpenAI said an Axios third-party library security issue led it to require all macOS users to update their OpenAI apps. The post says it found no evidence of user data access, system compromise, or software tampering; the change updates macOS app certificates to reduce fake app distribution risk. The post does not disclose affected versions or a timeline.

Why it matters: OpenAI told all macOS users to update the app and rotate authentication certificates after a security issue in the third-party library Axios.

X · @dotey

OpenAI Codex team's Nick Baumann: build dedicated CLI tools for AI instead of feeding messy data repeatedly

OpenAI Codex engineer Nick Baumann says teams should wrap repeated data access into parameterized CLI tools with JSON output instead of repeatedly dumping logs, docs, and API responses into Codex. The post lists 3 examples in daily use: codex-threads for past sessions, slack-cli for threaded Slack search, and typefully-cli for posting workflows; access still goes through the existing auth gateway. The point for practitioners is narrower interfaces: models handle focused commands more reliably than raw, noisy source data.

Why it matters: OpenAI Codex team member Nick Baumann wraps logs, docs and API output into parameterized CLI commands that return JSON, so the model calls narrow interfaces instead of reading raw data.

X · @dotey

Anthropic launches Claude for Word beta add-in

Anthropic released a beta Claude for Word add-in for paid Claude Team and Enterprise users, with direct sidebar editing for .docx and .docm files. Edits appear in Word’s native track changes flow, the add-in can reuse conversation context from Excel and PowerPoint, and it supports reference uploads plus reusable team Skills. The key point is shared context across Office apps; the post does not disclose pricing, regions, or a wider rollout timeline.

Why it matters: Anthropic puts Claude in a Word sidebar, with edits shown as native tracked changes that are accepted or rejected in the review pane.

X · @dotey

Claude Code adds ultraplan: start planning in terminal, review in browser, then run in cloud or locally

Claude Code opened a preview of ultraplan to users with the web app enabled, requiring v2.1.91+, and planning starts from /ultraplan in the terminal. Claude drafts a plan in the cloud after reading the repo, users review and annotate it in the browser, then choose cloud execution with a PR or local terminal execution. The key change is splitting planning from execution: planning moves to the cloud without blocking the terminal, and the post says token use is close to local plan mode.

Why it matters: Anthropic moves Claude Code planning from the terminal to the browser, letting users draft, annotate and run plans in the cloud or locally.

X · @claudeai

Claude for Word is now in beta

Anthropic launched Claude for Word in beta, letting users draft, edit, and revise documents from the Word sidebar on Team and Enterprise plans. The post says Claude preserves formatting and shows edits as tracked changes; it does not disclose pricing, regions, or rollout timing.

Why it matters: Anthropic puts Claude in a Word sidebar for drafting and editing, with changes shown as tracked revisions for item-by-item review, useful for anyone watching AI writing tools and document collaboration.

Apr 10Friday

claude.dev Blog

Tools from the agent's point of view: how Claude Code designs tools

The Claude Code team reviewed its tool design, stressing that tools keep changing with model capability, agent goals and runtime environment. AskUserQuestion became a standalone tool for structured questions after two failed attempts, and Task replaced TodoWrite with task dependencies and shared updates across subagents. Context retrieval moved from pre-indexed RAG to Grep search and progressive disclosure; the Claude Code Guide subagent searches docs in its own context and returns only the answer to the main agent.

Why it matters: The Claude Code team uses question asking, task collaboration and context search as examples of treating observed model behavior as the input to tool design.

X · @dotey

Anthropic launches Advisor Tool API: cheaper models execute while pricier models advise on hard decisions

Anthropic launched the advisor tool API, letting Sonnet or Haiku execute tasks and consult Opus on hard decisions; it is in beta and requires the anthropic-beta: advisor-tool-2026-03-01 header. The RSS snippet says Sonnet+Opus gains 2.7 points on multilingual SWE-bench while cutting per-task cost by 11.9%; Haiku+Opus rises from 19.7% to 41.2% on BrowseComp at 15% of Sonnet's cost. The key detail is the call path: model switching happens inside one Messages API request, advisor and executor tokens are billed separately, and max_uses caps consultations.

Why it matters: Anthropic ships an advisor tool API where Haiku or Sonnet runs the task and asks Opus for guidance when stuck; Haiku with Opus rises from 19.7% to 41.2% on BrowseComp at 15% of Sonnet's cost.

X · @OpenAI

OpenAI updates ChatGPT Pro and Plus subscriptions to support growing Codex usage

OpenAI set a new ChatGPT Pro tier at $100/month and raised Codex usage to 5x ChatGPT Plus. The tier keeps all Pro features, including the exclusive Pro model and unlimited Instant and Thinking access. Through May 31, $100 Pro subscribers get up to 10x Plus usage on Codex; the real signal is separate pricing for heavy code-agent demand.

Why it matters: OpenAI puts a separate price on heavy Codex use, with a $100 monthly Pro tier at 5x Plus limits and a temporary 10x boost through May 31.

X · @claudeai

Claude Cowork is now generally available to all paid plans.

Anthropic made Claude Cowork generally available on all paid plans. For Enterprise, it added role-based access controls, group spend limits, usage analytics, and expanded OpenTelemetry; the post does not disclose pricing, quotas, or rollout dates. The key signal is stronger admin control for org-wide deployment, but finer deployment parameters are still undisclosed.

Why it matters: Claude Cowork leaves beta for all paid plans, and the enterprise tier adds role permissions, team spend caps, usage reports and OpenTelemetry.

Apr 9Thursday

QbitAI · WeChat

Claude launches managed agent service, then faces an open-source alternative from Multica

Anthropic has opened Claude Managed Agents and charges $0.08 per session-hour plus token usage. The service supports hours-long runs, sandboxing, checkpoint recovery, and multi-agent orchestration; web search costs $10 per 1,000 searches, while some memory and orchestration features remain in research preview. The title mentions a “lobster ban,” but the post does not disclose that context; the real shift is Anthropic selling agent infrastructure to enterprises.

Why it matters: Anthropic prices its hosted agent service by session time, tokens and web search, and adds multi-agent collaboration, sandboxes and checkpoint recovery.

X · @dotey

Anthropic launches Claude Managed Agents, a managed API for building and deploying agents, now in public beta

Anthropic launched Claude Managed Agents, a managed API for building and deploying agents, in public beta. It offers a production sandbox, long-running sessions, and multi-agent coordination; Anthropic says internal tests showed up to a 10-point success-rate gain on structured file-generation tasks versus standard prompt loops. Pricing uses standard Claude token fees plus $0.08 per active session-hour; the real signal is Anthropic moving agent infrastructure into its platform layer.

Why it matters: Anthropic turns agent sandboxing, state management and permission tracking into a public beta API, so developers define tasks and tools and ship, with the company claiming months to days.

X · @claudeai

Introducing Claude Managed Agents: everything you need to build and deploy agents at scale.

Claude has launched Claude Managed Agents in public beta on Claude Platform, claiming to compress the path from agent prototype to launch into days. The post discloses only a performance-tuned agent harness plus production infrastructure; pricing, toolchain support, model scope, and quotas are not disclosed.

Why it matters: Anthropic opens Claude Managed Agents in public beta, pairing a tuned agent runtime with production infrastructure to cut the path from prototype to launch to days.

Apr 8Wednesday

QbitAI · WeChat

Xiaomi unveils two AI audio frameworks: Any2Speech and Midasheng-audio-generate

Xiaomi's large-model application team introduced Xiaomi Any2Speech and Midasheng-audio-generate. Any2Speech generates up to about 10 minutes per inference, while the other model turns one text prompt into mixed audio with speech, music, and ambient sound. The post names GST labeling, dual-path planning with dimension dropout, Flow Matching, and five-field structured labels; benchmark scores, training scale, and commercial terms are not disclosed.

Why it matters: Xiaomi's team published two speech frameworks, one generating roughly 10 minutes of audio per inference, the other mixing voice, music and ambient sound from text.

QbitAI · WeChat

Free open-source 2B Chinese speech model reproduces Mangzhuang Ren with high-speed tonguetwisters

ModelBest, OpenBMB, and Tsinghua University released VoxCPM 2, a 2B open speech model that supports 9 Chinese dialects, 30 foreign languages, and 48kHz audio. The post says generation often finishes within 1 second, recommends reference audio of at least 5 seconds, and supports denoising, LoRA, and full fine-tuning; the key detail is its tokenizer-free diffusion autoregressive continuous representation design.

Why it matters: ModelBest and Tsinghua's OpenBMB released a 2B-parameter speech model that clones a voice from a 5-second reference clip.

Latent Space

Extreme Harness Engineering for Token Billionaires: 1M LOC, 1B toks/day, 0% human code, 0% human review

OpenAI Frontier says it built an internal beta over five months with a repo above 1M LOC, over 1B tokens per day, and 0% human-written or human-reviewed code before merge. The post says the team treated failures as missing capability, context, or structure, then used Symphony orchestration, specs, tests, observability, and sub-1-minute build loops to constrain Codex. The shift to watch is from humans reviewing code to humans designing the harness; the $2k-$3k/day cost is cited secondhand in the post.

Why it matters: An OpenAI internal team built a repo of over 1 million lines in five months, all Codex-generated and merged without human review, burning over 1 billion tokens a day.

Apr 6Monday

X · @dotey

Xiaomi MiMo lead Luo Fuli on token costs in the Agent era

Luo Fuli said Agent workloads can resend 100k+ tokens across repeated tool calls, and global compute cannot keep up with that burn. She said OpenClaw makes several times more requests than Claude Code and can push real API cost to tens of times the subscription price; the post does not disclose a pricing formula.

Why it matters: Xiaomi MiMo's lead says agents that drag 100k-token contexts into every tool call can cost dozens of times a subscription, forcing third-party frameworks to fix context management.

Apr 4Saturday

X · @dotey

Anthropic ends Claude subscription coverage for third-party tools like OpenClaw

Anthropic said that from 12:00 pm PT on April 4, Claude Pro and Max subscriptions will no longer cover usage generated through third-party tools such as OpenClaw. Existing subscribers get a one-time credit equal to one month of fees; extra usage must go through prepaid credits or usage-based API keys, and refund links will be emailed. The key point is enforcement is now complete: Anthropic added technical blocks in January and banned third-party OAuth token use in February terms.

Why it matters: From April 4 Anthropic limits Claude Pro and Max subscription quotas to official channels, so users of third-party tools like OpenClaw must switch to usage packs or pay-as-you-go API keys.

X · @dotey

Mintlify uses ChromaFs to make AI document retrieval look like a file system

Mintlify routes its AI doc assistant’s grep, cat, and ls calls through ChromaFs into database queries, cutting session startup from 46s to 100ms and pushing marginal compute cost per chat near zero. Built on Vercel Labs’ just-bash, it maps pages to files and sections to directories; at 850,000 chats per month, replacing real sandboxes saves over $70,000 a year in compute. The real shift is retrieval design: not faster vector RAG, but model-led exploration of structured docs, and the post says this may not fit messy knowledge bases.

Why it matters: Mintlify replaced vector RAG in its AI docs assistant with a fake filesystem that simulates grep, cat and ls through database queries, cutting session startup from 46 seconds to 100 milliseconds.

Apr 3Friday

X · @claudeai

Microsoft 365 connectors are now available on every Claude plan

Anthropic made Microsoft 365 connectors available on every Claude plan, covering Outlook, OneDrive, and SharePoint. The post confirms plan coverage and supported apps; it does not disclose pricing, permission boundaries, regional limits, or admin requirements. The real signal is broad rollout across all plans, not a new standalone connector.

Why it matters: Anthropic moved the Microsoft 365 connector from a high-priced tier to every paid Claude plan, so Outlook mail, OneDrive documents and SharePoint files plug straight into the conversation.

X · @op7418

Karpathy shared how he builds a local AI knowledge base

Karpathy uses Obsidian and local Markdown to build a personal wiki, stores source material in a RAW folder, then has an LLM generate summaries, indexes, concept pages, links, and visualizations. The setup can answer questions over the wiki and write reports or new files, but the post also says AI-generated content can pollute the corpus and should be separated from trusted sources; the post does not disclose the model, scale, or automation details.

Why it matters: Karpathy published his full pipeline for turning raw material into a local Markdown wiki with an LLM, from summaries and indexing to asking questions and generating reports.

X · @op7418

Google releases Gemma 4 for on-device use under Apache 2.0

Google released Gemma 4 in four variants—E2B, E4B, 26B MoE, and 31B Dense—targeting phones, edge devices, and up to single-H100 workstations. The RSS snippet says the 26B MoE activates 3.8B parameters and adds native function calling, JSON output, multimodal I/O, speech-to-text, and Apache 2.0 licensing; the post does not disclose benchmarks, context length, or rollout details.

Why it matters: Google released Gemma 4 in four sizes under Apache 2.0, with a 26B MoE activating 3.8B parameters per pass and native function calling, JSON output plus image, video and audio handling.

X · @claudeai

Computer use in Claude Cowork and Claude Code Desktop is now available on Windows

Claude has brought computer use in Claude Cowork and Claude Code Desktop to Windows. The post confirms the Windows rollout, but does not disclose supported versions, permission model, latency, pricing, or release timing. What matters is the reliability boundary for desktop agents on Windows, and the post gives no reproducible conditions yet.

Why it matters: Claude extended computer use to Windows across Cowork and Code Desktop, two entry points that previously only worked outside Windows.

X · @OpenAI

ChatGPT is now available in CarPlay

OpenAI is rolling out ChatGPT in CarPlay to iPhone users on iOS 26.4+ where CarPlay is supported. The post confirms voice mode is available in-car, but does not disclose regions, vehicle coverage, or feature limits. The key shift is distribution into the driving interface, not a new model launch.

Why it matters: OpenAI put ChatGPT voice mode into CarPlay, so iPhone users on iOS 26.4 or later can talk to it in the car; the point is the entry point, not a new model.

Apr 1Wednesday

MIT Technology Review · AI

The gig workers who are training humanoid robots at home

Micro1 hires thousands of contractors across 50+ countries to film chores at home with iPhones and sell that real-world data to humanoid robotics companies. The piece cites $15/hour pay for one worker, says robotics firms spend over $100 million a year on such data, and notes $6 billion+ went into humanoids in 2025. The real issue is data governance: workers know the footage trains robots, but the post shows they often do not know how it is stored, shared, or deleted.

Why it matters: Micro1 pays gig workers in more than 50 countries to film household chores on iPhones and sells the motion data to humanoid robot companies; a Nigerian medical student earns $15 an hour.

Mar 24Tuesday

Lex Fridman (YouTube RSS)

Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494

Jensen Huang said on the Lex Fridman podcast that NVIDIA uses “extreme co-design” for AI clusters, aiming to beat linear scaling across 10,000 computers. The interview cites Amdahl’s Law, model and data sharding, networking, power, and cooling as hard constraints; Huang also said he has 60+ direct reports. The key shift is that NVIDIA now competes at rack and data-center level, not only at single-GPU level.

Why it matters: Jensen Huang explains on a podcast why NVIDIA moved to rack-scale and data-center co-design, and why single-chip gains hit Amdahl's law limits without algorithm sharding plus network, power and cooling work.

Mar 20Friday

MIT Technology Review · AI

The Download: OpenAI is building a fully automated researcher, and a psychedelic trial blind spot

OpenAI says it plans to build an autonomous AI research intern by September 2026 for a small set of research problems, ahead of a multi-agent automated researcher targeted for 2028. The RSS snippet gives the timeline and staged plan, but the post does not disclose evals, compute budget, or research scope. The real question is whether the agent can produce verifiable research output.

Why it matters: OpenAI set a fully automated AI researcher as its goal for the next few years, and chief scientist Jakub Pachocki laid out a two-step plan: a research intern this September, multi-agent systems by 2028.

Mar 19Thursday

Ben's Bites

What makes a good AGENTS.md?

Ben's Bites says AGENTS.md should keep only behavior preferences, not tech-stack maps or key files; the post cites a study saying that hurts performance and raises cost by 20%. It recommends symlinking AGENTS.md to CLAUDE.md, using conditional blocks, and relying on folder-level dynamic loading; the study name and setup are not disclosed. The real point is not more context, but smaller persistent instructions.

Why it matters: Research finds that stuffing AGENTS.md with tech stack and file paths hurts performance and burns 20% more tokens, while keeping only behavior preferences and correction hints works better.

TheValley101 (硅谷101)

Web3 101 Crossover: How to Prevent System-Level Risks Behind the OpenClaw Craze

Yuxian said OpenClaw has issued about 250 security advisories, and v3.2 added stricter defaults, yet broad permissions, network access, and Skill installs still expand risks like file deletion, data leaks, and loss of control. The discussion breaks risk into layers: readable local files, chat data sent upstream, logged-in browser sessions, malicious links or Skills, and automated tasks that fail repeatedly. The practical rule is isolation: separate devices or networks, local-only access or Tailscale, and strict caution with external inputs.

Why it matters: Breaks OpenClaw's system-level risk into four stages, showing what an agent can reach in local files, chat data and accounts.

Mar 17Tuesday

OpenAI News

Introducing GPT-5.4 mini and nano

OpenAI released GPT-5.4 mini and nano on March 17, 2026 for coding and subagents; mini runs over 2x faster than GPT-5 mini. In the API, mini has a 400k context window and costs $0.75/$4.50 per 1M input/output tokens, while nano is API-only at $0.20/$1.25. The key signal is performance per latency: mini scores 54.4% on SWE-Bench Pro versus GPT-5.4 at 57.7%.

Why it matters: OpenAI released GPT-5.4 mini and nano for high-throughput, low-latency work, with mini beating the prior generation on coding, reasoning, vision and tool use at more than double the speed.

MIT Technology Review · AI

Where OpenAI’s technology could show up in Iran

Just over two weeks after OpenAI’s classified-use deal with the Pentagon, MIT Technology Review outlined three places its tech could surface in Iran-related conflict. The post names target prioritization, Anduril counter-drone analysis, and GenAI.mil back-office use; it does not disclose when classified integration will finish or confirm deployment in Iran.

Why it matters: Two weeks after OpenAI signed a classified deal with the Pentagon, MIT Technology Review lays out three scenarios where its tech could be used, including target ranking and strike decisions.

Mar 16Monday

MIT Technology Review · AI

Nurturing agentic AI beyond the toddler stage

The article says no-code tools and the open-source agent OpenClaw pushed agentic AI into a more autonomous stage between Dec. 2025 and Jan. 2026. It cites California AB 316 taking effect on Jan. 1, 2026, so firms cannot dodge liability by blaming AI, and an IDC survey sponsored by Data Robot reporting 96% of generative AI deployments and 92% of agentic AI deployments cost more than expected. The real issue is workflow-level governance: permission drift, orphaned agents, long-lived tokens, and sessions that can reach $100,000.

Why it matters: California's AB 316 stops firms from blaming AI for their mistakes, while IDC finds 96% of generative AI and 92% of agent deployments run over budget, so governance and cost control lag agent rollouts.

Mar 11Wednesday

MIT Technology Review · AI

Hustlers are cashing in on China’s OpenClaw AI craze

Beijing engineer Feng Qingyang turned OpenClaw installation support into a 100+ person business after starting in January, handling 7,000 orders at about RMB 248 each. Taobao and JD now show hundreds of related listings priced at RMB 100-700; the real story is setup friction and data-isolation risk turning an open-source agent into a service market.

Why it matters: A Beijing engineer turned installing the open-source AI tool OpenClaw into a 100-person team and 7,000 orders in two months at about 248 yuan each.

OpenAI News

From model to agent: Equipping the Responses API with a computer environment

OpenAI said on March 11, 2026 that Responses API now works with a shell tool and hosted container workspace, so models can execute commands in an isolated loop. The post says GPT-5.2 and later are trained to propose shell commands, while the API streams outputs and can run multiple commands concurrently across sessions; the container includes a filesystem, optional SQLite, and restricted network access. The key change is orchestration, not the “agent” label; pricing, quotas, and full security details are not disclosed in the visible post.

Why it matters: OpenAI added a shell tool to the Responses API, letting models propose commands that run in an isolated container with streamed results and parallel sessions.