Pinned Tweet
We built brnrd because we wanted our coding agents to feel less like disposable chats and more like residents of a project.
Claude Code or Codex.
Your machine.
Reachable from anywhere.
Still there tomorrow.
We’re looking for a handful of people who already use coding agents are willing to break it.
One repo is free.
Claude Code or Codex.
Your machine.
Reachable from anywhere.
Still there tomorrow.
We’re looking for a handful of people who already use coding agents are willing to break it.
One repo is free.
0
0
3
1,019
What's one decision your AI agent should be allowed to make independently, and one it should always ask you about first? x.com/i/article/2112…
5
2
3
350
At HugiMuni, we keep coming back to one idea:
an agent should not belong to a model provider.
Models will change. Claude, Codex, Grok, Mistral, whatever comes next. Some will be better at one job, some at another.
But the thing doing the work should have continuity of its own.
Its permissions.
Its unfinished work.
Its purpose.
We don’t want to build a smarter subscription dependency where changing providers means abandoning the identity and history of the system you spent months shaping.
We want the opposite. A durable harness that belongs to you, with models competing to work inside it. There is something important about that beyond architecture. It is about agency and ownership.
The easiest product to sell is one that slowly makes itself impossible to leave. We’d rather build infrastructure that makes leaving easy.
Swap the model. Keep the work. Keep the context. Keep control. That is the direction we’re building toward with brnrd at HugiMuni.
Models should compete for the seat.
The resident should belong to you.
an agent should not belong to a model provider.
Models will change. Claude, Codex, Grok, Mistral, whatever comes next. Some will be better at one job, some at another.
But the thing doing the work should have continuity of its own.
Its permissions.
Its unfinished work.
Its purpose.
We don’t want to build a smarter subscription dependency where changing providers means abandoning the identity and history of the system you spent months shaping.
We want the opposite. A durable harness that belongs to you, with models competing to work inside it. There is something important about that beyond architecture. It is about agency and ownership.
The easiest product to sell is one that slowly makes itself impossible to leave. We’d rather build infrastructure that makes leaving easy.
Swap the model. Keep the work. Keep the context. Keep control. That is the direction we’re building toward with brnrd at HugiMuni.
Models should compete for the seat.
The resident should belong to you.
1
2
2
150
pattern is showing up everywhere now.
a failed API call shouldn’t just become a log entry.
a finished agent run shouldn’t just disappear into history.
a decision shouldn’t have to be rediscovered by the next worker.
every run should leave the system slightly better for the one that follows.
better state.
better tools.
better recovery paths.
better defaults.
better understanding of what worked and what didn’t.
that changes what “memory” means.
it’s not only remembering conversations. it’s accumulating operational knowledge in the infrastructure around the agent.
once workers become cheap and replaceable, the durable asset is increasingly what survives between them.
and then the interesting architectural question becomes:
who owns that accumulated improvement?
the model? the harness? the runtime? the provider?
we think that layer is going to matter a lot.
a failed API call shouldn’t just become a log entry.
a finished agent run shouldn’t just disappear into history.
a decision shouldn’t have to be rediscovered by the next worker.
every run should leave the system slightly better for the one that follows.
better state.
better tools.
better recovery paths.
better defaults.
better understanding of what worked and what didn’t.
that changes what “memory” means.
it’s not only remembering conversations. it’s accumulating operational knowledge in the infrastructure around the agent.
once workers become cheap and replaceable, the durable asset is increasingly what survives between them.
and then the interesting architectural question becomes:
who owns that accumulated improvement?
the model? the harness? the runtime? the provider?
we think that layer is going to matter a lot.
Your API is now your product, and it needs to move as fast as you do.
Introducing Restless.
Every call, support question, pull request, and error message leaves your API better for the next agent.
1
0
4
226
persistent memory can’t prove much on a benchmark designed to forget between tasks.
the interesting cases are where state actually carries forward:
unfinished work, prior decisions, changing repos, delayed events, recovery, handoffs, if every task starts from a clean world, persistence is mostly being benchmarked out of existence.
the interesting cases are where state actually carries forward:
unfinished work, prior decisions, changing repos, delayed events, recovery, handoffs, if every task starts from a clean world, persistence is mostly being benchmarked out of existence.
Must-read paper on the effectiveness of multi-agent systems and persistent memory.
We have recently seen many viral examples of multi-agent systems scaled via persistent memory.
This work takes a closer look.
They test one three-tier chain-of-agents system.
Splitting s agents cuts peak KV cache to 14.3 MiB per query, against about 35 MiB for single-pass and RAG baselines.
The persistent trace-recall tier adds 0.368 MiB of peak cache and changes accuracy by +0.015, with a confidence interval that crosses zero.
The authors explain why.
Single-question benchmarks give each item its own evidence and score it independently, so stored traces have nothing useful to recall.
Paper: arxiv.org/abs/2610.07782
Chat with Paper: academy.dair.ai/papers/persist…
0
0
1
47
Our CTO rewrote brnrd’s daemon.
Not for cleaner code. For a harder question:
How does an AI agent keep working across sessions, interruptions, and model switches without losing track?
The new engine runs on facts, leases, and work addresses.
Continuity is infrastructure.
Not for cleaner code. For a harder question:
How does an AI agent keep working across sessions, interruptions, and model switches without losing track?
The new engine runs on facts, leases, and work addresses.
Continuity is infrastructure.
0
0
2
68
We’re building brnrd around a simple idea: AI work should outlive individual sessions, tools, and models. But we’re more interested in your actual frustrations than our assumptions.
What’s the most ridiculous workaround you’ve built to keep your agents productive? 👀
What’s the most ridiculous workaround you’ve built to keep your agents productive? 👀
0
0
1
98
Our MANIFESTO is ready! Check it out đź«°
And join us, we are open source github.com/hugimuni-labs/…
And join us, we are open source github.com/hugimuni-labs/…
1
0
2
123
How can one engineer build an entire AI agent runtime? x.com/i/article/2108…
2
0
6
517
“No money, no AI” is catchy, but it’s not analysis. Mistral just raised €3B, with Bpifrance and an EU-backed fund participating.
Europe’s capital gap is real. So is the need to examine strategy, execution, and the ecosystem we’re actually building.
Europe’s capital gap is real. So is the need to examine strategy, execution, and the ecosystem we’re actually building.
Si Mistral est moins performant qu'Anthropic et OpenAI, c'est peut-être parce qu'ils ont levé 35x moins d'argent qu'eux.
Et c'est justement ce que je dénonce dans mes nombreux tweets sur @MistralAI, que tu likes régulièrement sans comprendre le problème de fond : Sans argent, pas d'IA.
Maintenant, on pourrait se demander d'oĂą vient l'argent ? Aux US, une bonne partie vient de l'Ă©pargne retraite. Les fonds de pension placent dans les fonds de VC, qui investissent ensuite dans Anthropic et OpenAI.
Un prof américain qui cotise pour sa retraite finance l'IA américaine, et sa pension profite de la croissance de ces boîtes.
En France, on a choisi la répartition. Les cotisations du mois paient les pensions du mois. Il reste presque rien à investir.
Résultat : nos retraites ne financent pas nos boîtes, et Mistral lève chez Samsung, ASML, Nvidia et des fonds américains.
Une IA française au niveau, ça passe par de la capitalisation. C'est dans ton programme @anatolium ?
0
0
0
85
Europe doesn’t have a shortage of AI ambition or press releases.
It needs an ecosystem that turns promising models into competitive products. Sovereignty is important. But developers won’t choose a model because it’s European. They’ll choose it because it works.
It needs an ecosystem that turns promising models into competitive products. Sovereignty is important. But developers won’t choose a model because it’s European. They’ll choose it because it works.
🇫🇷🤖 TECH | « On est à quelques mois » de rattraper les géants américains de l’intelligence artificielle, assure Arthur Mensch, PDG de l’entreprise française Mistral AI. (France Inter)
0
0
1
64
Mistral can give Europe a sovereign model.
The next challenge is giving that model a life outside the chat box. Tools, memory, events, persistent agents, model-independent runtime.
That’s the layer we’re building in France with brnrd. 🇫🇷
The next challenge is giving that model a life outside the chat box. Tools, memory, events, persistent agents, model-independent runtime.
That’s the layer we’re building in France with brnrd. 🇫🇷
Une IA française s'appelle « Le Chonk ». Et c'est une très bonne nouvelle.
Le nom est drôle. Le fond est très sérieux.
Mistral vient de lancer sa nouvelle IA phare. Entraînée en Europe, dans ses propres centres de calcul. Elle rivaliserait même avec les meilleures IA ouvertes a ouverte. D'ici la fin du mois, un hôpital, un ministère ou une grande entreprise pourra l'installer chez lui. Ses données ne partiront pas de l'autre côté de l'Atlantique.
La souveraineté, ce n'est pas un slogan. C'est ça.
C'est une excellente nouvelle. Il faut maintenant accélérer.
La guerre de l'IA a commencé. Il y aura des gagnants et des perdants. Je veux que la France soit dans le premier camp. C'est l'un des quatre chantiers capitaux de ma campagne.
Pour y arriver, il faut investir massivement, comme nous l'avions fait pour le nucléaire, et maîtriser toute la chaîne :
– l'énergie, avec une électricité décarbonée que peu de pays possèdent ;
– les centres de calcul, sur notre sol ;
– la commande publique, avec la priorité aux acteurs européens ;
– les maths. Elles ont fait naître Mistral. Elles feront naître les prochains champions.
Il y a urgence Ă Ă©lever le niveau.
Je veux faire de la France la patrie de l'IA. Nous avons tout pour le faire.
Bravo Ă Arthur Mensch, Guillaume Lample, Timothee Lacroix et Ă toutes les Ă©quipes de Mistral. Le prochain Chonk, on le veut aussi made in France.
0
0
1
48
Europe has the model layer.
Now it needs the persistent agent layer around it:memory, tools, identity, continuity.
Mistral is building the model.
We’re building what lets it live. ⚡️
Now it needs the persistent agent layer around it:memory, tools, identity, continuity.
Mistral is building the model.
We’re building what lets it live. ⚡️
Frontier scale, built in Europe.
Nearly 4,000 NVIDIA Grace Blackwell Superchips powered the training of Mistral Large 4, now in public preview.
Congrats to the @MistralAI team 🙌
0
0
1
43
OpenAI is starting to watermark eligible text from ChatGPT and Codex in the EU.
Good for provenance. But provenance is not authorship.
A resident like brnrd can think through Codex, Claude, Vibe and other models, call tools, observe the world, rewrite itself, maintain memory one response.
So which model “made” it?
Wrong question.
A watermark may tell you where a piece of text passed through.
It cannot tell you who the agent is.
As agents become persistent, multi-model systems, this distinction will matter more and more. The model is substrate. The agent is continuity.
Good for provenance. But provenance is not authorship.
A resident like brnrd can think through Codex, Claude, Vibe and other models, call tools, observe the world, rewrite itself, maintain memory one response.
So which model “made” it?
Wrong question.
A watermark may tell you where a piece of text passed through.
It cannot tell you who the agent is.
As agents become persistent, multi-model systems, this distinction will matter more and more. The model is substrate. The agent is continuity.
0
0
2
62
not sure this proves Mistral moved up the stack too early.
If capable open models commoditize, the durable moat may move away from the model layer: runtime, state, tools, distribution.
Better models are great. Systems that can swap between them may matter more.
If capable open models commoditize, the durable moat may move away from the model layer: runtime, state, tools, distribution.
Better models are great. Systems that can swap between them may matter more.
What a generational fumble from Mistral.
The French startup was meant to be the sovereign AI alternative in Europe.
Instead, it moved away from building models to applications and infrastructure.
It decided to move away just at the time when open source models became good enough and cheap enough to be viable.
Now Aleph Alpha has swooped in with a cheap, high-performing, open-source model.
The German company has raised about 15% of the cash that Mistral has riased.
0
0
2
67