Anthropic filed to go public at a valuation that could top $1.4 trillion, revenue up 14x in a year, first operating profit last quarter. Then this detail: Claude wrote over 90% of Anthropic's own code in September. At what point do you stop reviewing the diffs? cnbc.com/2026/11/10/anthropi… #AI
1
2
48
Brad Trnavsky retweeted
ABOUT 100K PEOPLE ARE WAITING ON THE HERMES MOBILE APP. 👀
Here's what we know:
▸ Full release is aimed for after Thanksgiving and before Christmas
▸ The beta is capped by Apple and Android tester limits. The first 1K seats went to the people at NousCon
▸ More tester seats should open soon, and they go out from the waitlist
▸ It's not in the App Store or the Play Store yet
▸ It connects to your Hermes Cloud agent or your own self-hosted Hermes, and you can drive the same session from your phone and your desktop
▸ You can sketch a drawing and send it to your agent with your prompt
Hermes in your pocket by Christmas. That's the target. 🎄 🎅
Brad Trnavsky retweeted
I asked fable 5.2 to invent a new musical form.
Fable 5.2 wrote a symphony that’s shaped like a training run.
The orchestra starts as random noise, learns the key, pulse, harmony and the melody, then plays music it was never shown.
This is really interesting. I’m running a 6 agent panel on it to review and figure out ways to work this research into my workflow. Take a look: From Code Foundation Models to Agents and Applications: A Comprehensive Survey and Practical Guide to Code Intelligence
This is huge.
A group of 50 AI researchers (ByteDance, Alibaba, Tencent + universities) just dropped a 303 page field guide on code models + coding agents.
And the takeaways are not what most people assume.
Here are the highlights I’m thinking about (as someone who lives in Python + agents):
1
1
27
If this is accurate, the West will continue to fall behind in STEM fields and lose its competitive edge in business in general. I do not understand why 66% of Americans are stuck on AI bad? We watch too many movies here and listen to people with uneducated opinions.
44
Brad Trnavsky retweeted
Harvard now teaches AI agents all the way down to the GPU
the course is CS2680, Modern AI Systems, and it starts where most tutorials stop
the top of the stack comes first:
> Claude Code and how it runs
> the agent loop
> evals
> what every token costs
then it goes down one layer at a time:
> model deployment
> routing
> KV cache
> quantization
> speculative decoding
the clever part is the homework. all 5 assignments take apart one agent system: first you study how it works, then you build its harness yourself, then you optimize it down to the inference layer
it answers the question most AI courses skip: how does a modern agent get from the app you type into to the hardware it runs on
if you want a job in agent infra, this is the syllabus to copy, and the harness is where every assignment starts
the loops and graphs that sit on top of that stack are below ↓
Brad Trnavsky retweeted
90% confidence from an AI system can still lead to a wrong decision.
More data and stronger models do not remove the need for human judgment.
The ability to question a recommendation may become as important as the ability to use the technology.
Are we becoming better at thinking with AI, or simply more comfortable accepting its answers?
AI adoption may be inevitable. AI valuations are not.
When an industry requires trillions in capex, hundreds of billions in new financing, and increasingly complex debt structures, the key question isn't growth. It's sustainability. Can we keep this growth rate up? Will the funding hold? Is the new political environment going to shut it down? All valuable questions right now...
26
I have learned not to trust a single index score, so here are two. Google's Gemini 4 Carbon sits at number 1 on the Artificial Analysis Intelligence Index with a 59, and independent tracker Opper agrees at 60. It costs $7.76 per task and uses 29% fewer output tokens than Claude. Token efficiency is turning into the real contest. Which number do you watch first, cost or tokens? freakyjolly.com/gemini-4-car… #Gemini #AI
1
1
56
Brad Trnavsky retweeted
Hermes can now sign into websites using the logins you've already saved in your browser. 👀
69 PRs merged on November 7.
5 changes worth knowing:
▸ The in-app browser can now reuse signed-in sessions from your real browser profile, or fill logins from Hermes Vault, Bitwarden, and 1Password
▸ Saved Chrome and Brave passwords can now appear as a read-only login source when real-profile access is enabled
▸ Desktop and Webapp now share saved gateway connections on the same machine. The Webapp can switch between gateways and open remote SSH terminals
▸ Every new Group Chat runs on a gateway, including teams with named bot profiles on other computers. They keep working even after Desktop closes
▸ Gateway clients can now read saved conversations and subagent transcripts without opening the sessions, including their compaction history
Luke has the full 69 below 👇
Hermes merged 69 PRs on November 7
The ones worth knowing:
- Sign in to sites in the in-app browser with saved logins and your real browser profile
- Saved passwords from your real browser profile can fill logins as a read-only source
- Every new group chat runs on a gateway, and named profiles on other computers can join
- Desktop and the Webapp share one list of saved gateways, and the Webapp can reach them all
- Group chats no longer drop a message sent while the previous one is still settling
- Group bots no longer stay stuck or silent behind an abandoned turn
- Yuanbao leaves core for a Hermes Labs plugin, and existing users migrate automatically
- The Webapp can test saved SSH gateway connections and open saved SSH terminals over SSH
- Clients that paste images can use `hermes clipboard image --json`
- You can read a stored session without attaching, including picker fields, transcript pages, and compactions
- llm.oneshot drops schema and redact; plugins get structured output through their own RPC method
- Each delegate card shows only its own subagents
Here is the full list:
DESKTOP IMPROVEMENTS (13 PRs)
- feat(desktop): sign in to sites in the in-app browser with saved logins and your real browser profile
- One list of saved gateways for Desktop and the Webapp; the Webapp reaches them all
- fix(webapp): open saved SSH terminals over SSH
- fix(webapp): test saved SSH gateway connections
- Webapp Test and SSH Connect share one error-status table
- A removed saved gateway stays removed after restart (Desktop and Webapp)
- fix(desktop): each delegate card shows only its own subagents
- fix(desktop): /model X --provider Y switches through the canonical model change
- fix(desktop): persist Bot Chat wake diagnostics
- fix(desktop): install get-windows on macOS and Linux again
- fix(desktop): toast the message of an object-shaped error detail
- fix(tui): a refused ticket mint shows its repair command again
- docs(browser): local engine is a config choice, not a fallback rung
DEVELOPER TOOLS & PLUGINS (23 PRs)
- feat(cli): add `hermes clipboard image --json` for clients that paste images
- llm.oneshot drops schema and redact again; plugins get structured output through their own RPC method
- feat(gateway): read a stored session without attaching (picker fields, transcript pages, and compactions)
- feat(gateway): the control socket answers which gateway serves a profile
- feat(gateway): a client can name itself on session.create
- feat(shared): the gateway client takes its session source and client name
- fix(gateway): plugin RPC is a detectable contract member and session events work hosted
- fix(plugins): one plugin per session env name, and names stay in their own profile
- feat(memory): pass providers the previous user message on turn start
- chore(plugin-catalog): re-pin advanced-delegation to 13b2c1c (1.1.1)
- docs(agents): route core and plugin work to the boundary lint rules, and trim root to make room
- refactor(pm): keep the pnpm pin only in pm/lock.json
- fix(pm): one npmrc reader and file rule for the registry lookup and the build bridge
- fix(pm): a packaging change rebuilds the dependency environment
- fix(pm): pin pnpm's android-arm64 executable for linux-arm64-bionic
- test(js): refuse a re-pin before a completed install is reused
- chore(js): pnpm switch leftovers (cache inputs, undeclared deps, lock attributes, Nix hash docs)
- fix(js): refuse npm installs in the JS workspace folders, keep npx and npm run
- fix(ci): CI guards fail when they cannot see a change
- fix(tools): self-repo guard suggests the pruned scratch folder for clones
- refactor(paths): one owner for the install root, and a lint against new folder counting
- fix(acp): hermes-acp serves the selected profile
- test(acp), docs(shared): two review follow-ups after merge
SECURITY & SECRETS (3 PRs)
- feat(vault): real-profile saved passwords as a read-only browser: login source
- fix(left-core): administrator-pinned platform credentials outrank stale user extras
- fix(dashboard-auth): a foreign Portal JWT bearer is a 401, not 'Auth provider unreachable' 503
INTEGRATIONS & CHAT (9 PRs)
- Every new group chat runs on a gateway, and named profiles on other computers can join
- Group chats no longer drop a message sent while the previous one settles
- Group bots no longer stay stuck or silent behind an abandoned turn
- Yuanbao moves out of core into a Hermes Labs plugin; existing users migrate automatically
- Catalog: yuanbao Hermes Labs plugin, the new home for the Yuanbao platform leaving core
- fix(groups): complete restart stop recovery and keep controls responsive
- fix(groups): Disband asks a deferred turn's Stop at the home that admitted it
- fix(gateway): Daybreak delegation routes recognize the owner's gateway-hosted chats
- fix(delegation): a split unit's failed task keeps its own goal
STABILITY & FIXES (21 PRs)
- feat(compat): serve a moved module under its old names as one module object
- fix(update): refuse a target that lacks a startup file it declares
- fix(update): remove the folders a dropped package leaves behind
- fix(update): the downgrade guard finds state modules inside the package
- fix(update): the running updater loads no code from the tree it just moved
- fix(update): rewrite gateway launch files before any gateway starts
- test(compat): freeze the names v0.21.6+ updaters import after the swap
- test(upgrade): start from real releases; the next release moves a file
- fix(config): script and Docker installs keep the pre-update snapshot
- chore: delete code nothing calls
- fix: optional-import guards keep falling back after a module moves
- fix(logging): log routing keeps package-qualified logger names
- fix(cli): drop a dead config fallback and an impossible PROJECT_ROOT import
- fix(gateway): a held restart wait names its managed worker; the handoff cell caps it
- fix(profiles): a profile imported, installed or renamed onto a deleted name is live
- fix(gateway): two RPC failure paths log their error instead of dropping it
- fix(tui-gateway): a client's error answer to a prompt is logged at warning
- refactor(state)!: remove the CJK FTS5 tokenizer from core
- fix(state): a probe copy that cannot be reopened for its final scan is not damage
- fix(gateway): a profile store that moved with its home keeps its sessions
- fix(memory): previous_message is the typed ask after a multimodal turn
69 new features and fixes
nitter.cf/iamlukethedev/status/2…
I just pushed an update to storm-fusion-research-pipeline.
Based on Stanford's STORM research framework to handle multi-perspective synthesis without the manual grunt work. Packaged as a Hermes plugin, but it runs with any agent.
Grab it from the Hermes plugin repo or on my GitHub; test it out, and let me know what breaks. github.com/bmtrnavsky/storm-…
1
62
I just upgraded my local AI memory tonight.
Hindsight 1.2.0 + GPT OSS 120B + hs-decide-1-flash handling recall, consolidation, and reflect. Hs-decide-1-flash is now assisting in Recall, Reflect, and Consolidation, so the big LLM gets 31% fewer LLM calls.
Who else is running a persistent agent memory layer locally? What are you using? I have been VERY happy with #Hindsight. I hope these changes were worth it.
I'm thinking about using JEV to judge done, test completion, etc., at my gates in my Disciplined Vibe Coding and Spec-Driven Data Analysis Hermes plugins. Has anyone else tried this? Would you make it warn you, stop the process, or just log the failure?
4
285
I’m trying to decide if this is useful or just entertaining. Several tools do this I haven’t used any yet. Do you? Is it good? How does it help you.
This might be the most fun Hermes plugin yet.
You can now watch your Hermes agents work.
Office 3D turns every Hermes profile into a little person at a desk, in a live office right inside Hermes Desktop.
What it does:
▸ Agents doing work sit at their desk typing
▸ Free agents take a tea break in the lounge
▸ Offline agents leave an empty desk
▸ Click any desk to see its sessions, tokens and cost
▸ A leaderboard ranks your agents by tokens or by cost
▸ A status bar chip shows how many are busy at a glance
▸ Hermes can check the office too, so you can just ask who's working
What you can do with it:
▸ Run a research agent, a coding agent and a trip planner at once and see who's still busy without opening a single chat
▸ Spot which agent is eating your budget before the bill shows up
▸ Get a desktop ping when an agent finishes and heads to the lounge
▸ Give each part of a small business its own agent (support, social posts, bookkeeping) and watch the whole team in one view
Running a team of AI agents finally looks like running a team.
34
Shopify post-trained an open model for one Sidekick job, reading GraphQL schemas and docs, and it beat the frontier model on accuracy while answering 7x faster at 90% lower serving cost. This is the pattern I want for fixed agent jobs: own the small specialist, rent the frontier only for the planning. What is your most token-heavy agent job, and has anyone priced the specialist version? shopify.engineering/open-mod…
Brad Trnavsky retweeted
Hermes Agent users, bookmark this one.
@witcheer's Hermes Wingtips are now on HermesWatcher.
witcheer at Nous Research drops short, practical Hermes Agent tips. One feature or setting at a time. Now you can browse and search them right alongside the site's guides, skills and plugins, in witcheer's own words.
▸ All 113 Wingtips, 26 Autopilot prompts and 9 release explainers
▸ Every cited source, right on the card
▸ Search all of it as you type
▸ Open any tip in place, or share a direct link to it
▸ New tips show up every morning on their own
While you're there: install and chat app setup guides, a setup planner, and searchable skills and plugin directories. All free.
The fastest way to get good at Hermes.
hermeswatcher.com/wingtips
Brad Trnavsky retweeted
It is becoming f*cking dangerous
Anthropic just dropped one of the best tricks for Claude Code
The concept: run Sonnet, Haiku, and Opus 5.5 simultaneously, with each model handling only what it does best
Get Opus-level decision-making on every critical choice without paying Opus prices for routine operations
Simply launch Claude Code like this: claude --model sonnet --advisor opus
Here is the breakdown of responsibilities:
> Sonnet 5.5 runs the active session and writes the code
> Haiku 5.5 explores files and documentation in parallel
> Opus 5.5 stays on standby as an advisor for major decisions
> You stop wasting Opus budgets on tasks that don't require it
Opus only steps in during three specific scenarios:
> Before locking in a plan
> When the exact same error happens twice
> Before declaring the task complete
Sonnet builds. Haiku explores. Opus reviews
JEV engineering applies the exact same logic one layer down: micro-decisions that require no deep reasoning (choosing files, selecting tools, deciding to retry or stop) are processed by JEV in milliseconds, leaving the large models to handle only the actual splits
The complete tree structure:
Sonnet 5.5 runs the primary session and edits code search maps the repository using Haiku 5.5 grep finds call sites using Haiku 5.5 docs pulls API references using Haiku 5.5 Opus 5.5 acts as an on-call advisor, reading the full transcript when consulted JEV routes micro-forks within milliseconds
Paste this tree and prompt directly into Claude Code:
Rebuild my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 acting as an advisor
2. Create three read-only Haiku 5.5 subagents: search (maps the repository) grep (finds call sites), docs (pulls API references)
3. Add a single rule to CLAUDE.md: consult the advisor before locking a plan, after encountering the same error twice, and before marking a task as done
4. Route fast deterministic forks (tool selection, path routing, retries) through JEV. Show every change as a diff first. Do not apply edits until I give the go-ahead
code.claude.com/docs/en/advi…
Save this, you will use it every single day
Not gonna lie I LOVE the GOAT 🐐 plan. I plan and review with opus and build with Deepseek Flash. $100 in tokens for $10 is crazy good value.
GOAT just made $100 credits PERMANENT
On $10/mo 🐐 plan
DeepSeek V4.1 Flash 1017: $100 usage
GLM 5.3 Flash: $100 usage
GLM Flash casually doing 300-500 TPS in Command Code harness
full weight models, btw
2
55
Brad Trnavsky retweeted
I used to think this too until AI solved the Erdős conjecture. At this point, arguing that AI is simply synthesizing existing knowledge is like arguing that Einstein never really came up with General Relativity because he was simply synthesizing the work of Riemann, Poincare, etc.





















