Agent news
NEWS
Coverage of AI agent launches, model releases, pricing, and security, with links to related software and agent profiles.
Railway Says Customer Secrets May Have Been Decrypted. Customers Saw Abuse Hours Before the Warning.
The hosting platform told users to rotate every environment variable at 5:51 PM ET. Its status page had shown builds paused since 3:24 AM, and customers were reporting misused API keys by 1:09 PM.
Claude Duet is Anthropic’s proactive agent app, and it asks before it sends, pays or deletes
The research preview runs on Claude Opus 5.5 in its own cloud computer, checks your connected apps every few hours and shares Pro and Max usage limits. It opens by waitlist to US subscribers.
MCP reference servers 2.0 speak the stateless 2026-07-28 spec, and break a few things for clients
Eight days after 1.0, the official MCP servers move to the v2 SDKs and answer both 2026-07-28 and 2025-era clients from one process. The TypeScript servers now require Node 24, and error handling changes in ways clients may notice.
OpenClaw 2026.10.5 makes updates finish and conversations survive restarts
The new `latest` release is mostly a reliability cut: updates that recover on their own, turns and subagents that resume after a Gateway restart, and a lighter Gateway. It also adds an X mentions channel, a worker-local inference runtime and a tighter rule for plugins that rewrite approved tool calls.
Perplexity open-sources SPACE, the microVM sandbox behind Computer, with its red-team results attached
SPACE runs each agent in its own Firecracker microVM and keeps credentials and network policy outside it. Perplexity’s own testing found the VM boundary held, while the network rules needed fixing, a weakness it also reported in seven other sandbox platforms.
Vercel Drives reaches GA, giving sandboxed agents a workspace that outlives the sandbox
Persistent drives for Vercel Sandbox are now generally available on every plan. An agent’s files can follow it into a new sandbox, but only one sandbox can write to a drive at a time.
Meta’s Muse Spark Open puts an 816B agent model in public weights, with strings attached
Meta released downloadable weights for Muse Spark Open, a 42B-active mixture-of-experts model with a 1M-token context. Artificial Analysis scores it as the strongest US open-weight model it has tested, but behind the leading Chinese open-weight models, and the license adds attribution and a revenue threshold for API providers.
Grok Voice 3.0 tops the independent speech-to-speech index at the same $0.08 a minute
SpaceXAI’s new voice model now runs every agent that uses the grok-voice-latest alias. Artificial Analysis ranks it first overall, but the same leaderboard shows GPT-Live-1 ahead on agentic support calls and a Gemini model ahead on human preference.
Hermes Agent v0.22.0 lowers the door: no-key setup, app connectors and one-click local models
Nous Research’s Threshold release lets a new desktop user start without an API key, connects the agent to Gmail, Linear and Notion, and runs open models locally through llama.cpp. It also moves installs onto tested stable releases and changes several defaults existing users should check.
Meta’s Muse Wonder reaches the API at $4/$20, while Wonder Full stays inside Muse Code
Meta’s new frontier pair targets personal and professional agents. Muse Wonder is in preview on the Meta Model API with a 1M-token context; the stronger Muse Wonder Full is limited to higher Muse Code plans, which default to training on your content.
The X Subscription puts Cursor on a Grok plan, with Claude and GPT billed separately
SpaceXAI now sells Grok, Grok Bot, Cursor and X Premium as one subscription from $8 to $200 a month. New plans share one weekly Grok usage pool, but third-party models in Cursor sit outside it and bill at API list prices.
MCP reference servers reach 1.0, and fetch now blocks private addresses by default
The official MCP servers repository returns to semantic versioning with its last release on the 1.x SDK. The fetch server gains a guard against reaching internal hosts, filesystem writes change behavior, and outside pull requests are now closed.
Claude Haiku 5.5 brings cheaper tokens and adjustable effort to subagents
Anthropic’s small model adds a million-token context and adjustable effort. The cheapest rates stop at 100,000 prompt tokens, and its strongest benchmark scores do not represent the default setting.
OpenAI’s Decisions API enters public beta for fixed-answer agent routing
The new endpoint turns text and images into probabilities, categories or rubric scores. Its narrow output contract brings input-only billing to routing work, but evaluation and permissions remain the builder’s responsibility.
Mistral Large 4: Le Chonk previews enterprise agents before the weights arrive
Mistral's new flagship has credible coding, cyber and document results. Its public preview offers discounted tokens, but independent tests show why agent buyers should watch the whole task bill.
EmbeddingGemma 2 brings multimodal search to local AI agents
Google's 740M-parameter embedding model puts code, images, audio and video in one search space. The useful question is what developers can retrieve locally, and what they can afford to keep in the index.
Anthropic expands Claude cyber access with three permission tiers
Defense, Red Team and Specialized Access widen eligibility for advanced Claude models, but stronger credentials, misuse monitoring and safety-system restrictions still determine what teams can do.
AG-UI 1.0 gives agent apps a stable contract for streams and approvals
CopilotKit’s frozen specification covers subagents, multimodal tool results and interrupted runs. Developers still need to distinguish wire compatibility from a drop-in SDK upgrade.
Gemini 4 Argon starts with cyber defenders, not a public API
Google’s new frontier model promises a one-million-token output ceiling and long-horizon coding gains. The first access is tightly restricted; its advertised API price is for a later launch.
Holo4 uses one model across screens, code and APIs, but its two weights have different trade-offs
H Company's Holo4 models bring GUI, code and tool use into one agent stack. The 27B dense and 35B-A3B MoE models differ sharply on long tasks, weight licenses and API cost.
OpenClaw 2026.9.7 adds an Agents API harness; ChatGPT sign-in is a separate beta
OpenClaw 2026.9.7 brings an Agents API harness, a separate ChatGPT sign-in beta, and safer database rollback—with a Windows update caveat.
OpenAI Dots bring an always-on agent to ChatGPT, with approvals still in the loop
Launched at DevDay on September 29, Dots can keep projects moving between conversations. The harder question is which work they may do without asking you.
GPT-6.1 Sol narrows the gap to Astra, but the price claim has boundaries
OpenAI's new model reaches Work, Codex and the API before Chat. Its standard input and output rates are one-fifth of Astra's; the benchmark and caching picture is more complicated.
SpaceXAI opens shared Grok Team Bots in public beta
The new team layer puts common instructions and tools behind a Bot that colleagues can use separately or bring into Slack. Access is limited to Teams and Enterprise plans.