Twitxy Saved

Baptiste Jamin

Verified account This account is verified. Learn more

@baptistejamin

Co-Founder/CEO @crisp_im - Speaking about AI, LLMs & code. - Bootstrapping & coding since 2015 the future of customer service.
Nantes, France crisp.chat Joined June 2010
585 Following 2.5K Followers 4.4K Posts

Latest posts 5

Baptiste Jamin Verified account @baptistejamin
AI support agents are black boxes. Not anymore.

Hugo Observability: a full timeline for every AI conversation. Model used, tools & MCPs called, reasoning at each step, latency, errors.

Nothing hidden.
Crisp Verified account @crisp_chat Sep 10
We've been quietly shipping all summer long. Here is what we've been up to:

We’ve shipped a lot at Crisp this summer. ☀️

But if we had to summarize the direction in one sentence:

AI customer support is becoming more powerful — so the tools around it need to become more transparent, controllable and easier to operate.

A few highlights from our Summer 2026 Product Update:

🔍 See exactly what your AI is doing

With @hugoaiagent Observability, teams can now inspect the full timeline of an AI-powered conversation: which models were used, which tools were called, how the AI handled each step, how long it took and where things went wrong.

AI shouldn’t be a black box, especially when it’s talking to your customers.

🧠 More choice in the models powering Hugo

We’ve expanded the AI models available inside Crisp and improved our own models to deliver more reliable answers and fewer hallucinations.

👥 A better way to organize support teams

Operator Teams can now be created once and reused across mentions, routing and Sub-Inboxes, with memberships staying synchronized automatically.

🎯 Separate instructions for Hugo Copilot

Your customer-facing AI Agent and the Copilot helping your support team don’t necessarily need to behave the same way.

You can now configure dedicated instructions specifically for Copilot.

📊 Better visibility into human support activity

New messaging metrics make it easier to understand message volume, workload and activity by operator and message type.

And beyond those larger releases, we’ve also been working on the fundamentals:

→ A refreshed Crisp interface
→ A completely redesigned Dark Mode
→ Stronger bot and spam protection
→ Better multi-domain security
→ More control over email tracking and privacy
→ Redesigned Sub-Inboxes
→ Faster workflows and analytics
→ A ~30% smaller initial Chat Widget bundle
→ Improvements across routing, translations, integrations, SDKs and mobile apps

Individually, some of these changes are small.

Together, they’re part of the same goal:
Make AI-powered customer support easier to understand, safer to operate and genuinely pleasant for support teams to use every day.

Full Summer 2026 Product Update 👇
crisp.chat/en/blog/summer…
View quoted post on X
Baptiste Jamin Verified account @baptistejamin
GPT 5.6 Sol is really good and better than Claude Fable 5 in my opinion even on complex tasks.

It’s more token efficient, costs less per token, adding to that OpenAI provides better prompt caching.

It’s going to be my next workhorse.
3 replies 0 reposts 12 likes 429 views View on X
Baptiste Jamin Verified account @baptistejamin
Note to the "AI experts" who ask us for a Model Router:

A router that switches models mid-conversation breaks the cache, drops context, and costs MORE.

The best thing it can do is not route.

"But stateless / single-shot calls!" ->

That's ~10% of your spend. You're adding a routing layer. And a thousand new edge cases, to optimize the cheapest part of your bill.

We'll just pick the right model once. Keep your router for side projects.
Baptiste Jamin Verified account @baptistejamin
I appreciate @MistralAI, and I'm proud the EU has a company like this. As a founder pushing EU solutions at @crisp, I genuinely want them to win.

But how are crucial problems just... not getting fixed?
Their TS SDK has 146 stars, and nobody seems to handle PRs or fix issues.

Mistral now supports prompt caching, but on TypeScript? You literally can't use it.

Worse: we detected issues in their tool-calling parsing.
Some basic tool calls inject raw JSON directly into regular AI string messages because the API isn't parsing them properly.

This is a billion-dollar unicorn handling developer support like a 5-person SMB. Come on. The EU AI champion deserves better engineering hygiene than this.