What shipped on the platform, and every new release of the agent templates you can deploy.
A messaging channel that stops working is now turned off cleanly while the rest of your agent keeps running. Your credential and settings are kept safe so you can restore the channel, and the channels list tells you what happened and when.
Core v2026.8.27 · upstream notes.
If one of your messaging channels stops working, your agent now keeps running. The channel shows as disconnected with the reason, and everything else carries on.
Core v2026.8.27 · upstream notes.
Fixes an issue where signing in to a connected tool could drop you on a login page instead of finishing. The agent's sign-in page has also moved; old links still redirect, so there is nothing to do.
Core v2026.8.27 · upstream notes.
A better model picker with new models to choose from, web search that works before you add any key of your own, PDF and file attachments, and smarter memory and scheduling.
Also fixes an issue where a Telegram connection problem could stop the whole agent. Your agent now stays up and simply shows Telegram as disconnected until you reconnect it.
Core v2026.8.27. Upstream notes: v2026.8.13, v2026.8.16, v2026.8.16.2, v2026.8.18, v2026.8.19, v2026.8.27.
Search past conversations by word or phrase, follow long-running work as it happens, and pin interactive results to a session dashboard. Audio and video now stay attached to the conversation, and there are more model providers to choose from.
Your agent also builds up long-term memory in the background, on by default. This background work uses a little of your credit; you can turn it off in your agent's settings.
Core 2026.8.1. See the upstream changelog shipped in the package for the complete list.
Ships upstream core v2026.8.27 (was v2026.8.3).
This release rolls up six upstream Hermes releases that had stacked up behind a stuck CI gate: v2026.8.13, v2026.8.16, v2026.8.16.2, v2026.8.18, v2026.8.19 and v2026.8.27 (Hermes Agent v0.20.1 through v0.20.6).
What you get
/model picker, a Ctrl+P command palette and a richer /status. New models across the pickers (GLM-5.3-Flash, MiniMax M3 free, MiniMax H3 Max video).Full detail is in the upstream notes for each release, linked below.
Compatibility
Nothing in this release changes how your agent is configured, and no action is needed on your side. Your existing settings, chat pairings and connected channels carry over untouched. We verified before shipping that the pairing store location and file format, the chat allowlist rules, the session header and the provider list all behave exactly as before, and the full wrapper contract suite passes against the new core.
Core v2026.8.27. Upstream notes: v2026.8.13, v2026.8.16, v2026.8.16.2, v2026.8.18, v2026.8.19, v2026.8.27.
The Files tab is now a full workspace: browse, rename, move, copy, upload, download, delete with undo, share links, and open files in place, with Markdown, CSV tables and images rendered. Works on a phone too.
Fixes an issue where chatting on included credits could return an empty answer.
Fixes an issue for agents updated from an earlier version where Claude Code and Codex could be set up incorrectly. They now always use your own sign-in.
Included credits now run through OpenCode, with no login needed. Claude Code and Codex are for your own subscription: sign in with your own account, and your credential stays on your agent.
Your agent now starts with a ready-to-use project and the included model selected, so it answers out of the box. Also fixes an issue where work could be saved in a place that didn't survive an update.
A managed T3 Code coding agent you deploy in a click. Chat with it, pair it with the T3 Code app, connect GitHub so it can work on your repositories, and use included credits or sign in with your own Claude or ChatGPT subscription.
Ask "how many" and Nova reads the whole document and counts every match instead of estimating. If a section can't be read, it tells you.
Nova now reads a document right through and answers exactly, at a fraction of the previous cost. Also fixes several ways a reply could quietly end without an answer.
Hand Nova several independent questions and it researches them side by side, then writes you one answer. Your budget is divided up front so parallel work can't overspend. Off by default per team.
Give Nova a goal and it keeps working on it between conversations, off by default per team. Your spending limits apply to every paid step, and every stop tells you which limit it hit and how far it got.
Correct Nova twice on the same thing and it writes the lesson down. /refine makes a lesson stick immediately, and you can undo one so it's never re-learned.
There is now a blog and the page you are reading.
The blog carries four kinds of post — guides, tutorials, comparisons and announcements — and every one is also published as raw markdown for assistants that read the site directly.
Both surfaces have feeds. /feed.xml covers the blog; this page has its own at
/changelog/feed.xml for anyone who only wants release notes.
Portuguese routing is in place at /pt, with translations to follow.
Older parts of a long conversation now fold into a summary instead of falling off the end, with nothing deleted, and long-running turns finish much faster.
Maintenance and reliability improvements.
Follow along by RSS. The feed carries platform entries only.