CEO at Slite (YC18) | We build the first knowledge base on autopilot, in remote since 2016!

I'm the founder of Slite & Super (YCW18) 👉 Super.work 

Super.work turn all your company data into answers & AI agents for your whole team.

It comes with the recipes and tools for your team to get instant benefit from AI at work

Slite is your team dead simple yet crazy powerful knowledge base, since 2017 🔥

I'm fortunate to lead a diverse team of exceptional individuals spread across 10+ countries 🇮🇳 🇬🇷🇱🇹 🇦🇷 🇬🇧 🇳🇱 🇫🇷 🇳🇴 🇺🇸 🇪🇸 🇩🇪 🇨🇦 🇵🇹 . Our journey includes going through YC W18 batch, raising $15M+ in funding from top-tier investors like Spark Capital and Index Ventures, and serving tens of thousands of teams monthly.

This is my first multi-agentic workflow.
Of course, it makes the job easier but more so, it feels like the future.

Here's the setup.
I connected Claude Code to two MCPs.
1. Sanity, which is our CMS
2. Super, which holds our product knowledge, positioning, and messaging.

And here's what actually happens.
1. I ask Claude Code to QA a content page.
2. It pulls the full page from Sanity, every field, metadata, structure, all of it.
3. It calls Super as a fact-checking agent and cross-references the page against our actual product knowledge base.
4. It comes back with a plain language review of what's off, what's outdated, what's missing.
5. It pushes the approved fixes directly back to Sanity through the MCP.

The whole loop runs without me opening the CMS once.

Claude Code is an incredible general agent. 
It can orchestrate, plan, read files, execute tasks. But it's not purpose-built for anything. It doesn't know your product, your positioning, your company's source of truth.

When you pair a general orchestration agent with a purpose-built knowledge agent, the quality of output changes completely. The orchestrator handles the workflow. The specialist handles the judgment.

That's the pattern I think we'll see everywhere. Not one super-agent that does it all, but multiple agents that are good at different things, working together.

Christophe Pasquier

CEO Slite & Super (W18) | Most accurate search & workflows on team data

A semi-viral tweet was praising Linear's Slack agent last week
Jori Lallo quote-tweeted it with one line that's been stuck in my head since.
"Linear agent is soon just Linear."

He went on to explain that the agent will eventually have full feature parity with the UI. Everything you can do in the product, the agent can do too.

If you sit with that for a second, it reframes what a software product even is.

The UI becomes one client among many. Slack is a client. Claude Code is a client. A script someone wrote on a Sunday afternoon is a client. They all talk to the same system underneath, and they're all equally valid ways to use the product.

Most SaaS companies were built the other way around. The interface was the product, the API was an afterthought you shipped because enterprise customers kept asking. Jori is describing a world where that's completely inverted, where the system of record is the product and every interface, including the one with pixels, is just a client.

We're building toward the same thing at Slite. We shipped our MCP recently and we're still expanding it, exposing more endpoints, making more of the product fully programmable through agents. The vision is to turn Slite into a collaborative Obsidian for company knowledge, where agents search, update, reorganize, and verify docs just as naturally as a teammate would.

All SaaS products built on data as a moat are an API now, whether they like it or not.

Folks! Wanna catch up irl in Berlin next week?
I'm speaking at MCP Connect on March 3
And I'm genuinely excited for 2 reasons.

It’s college friends organising this, the founders behind Alpic, they’re moving fast on being the reference platform for anyone building MCPs and apps in the GenAI majors and it’s a blast to share stage with them. 

And ofc - Viktor Zamaruiev, Director of Engineering at Superhuman, is also speaking!

He's giving a talk on how they're thinking about AI at Superhuman. I've been a Superhuman user for years, and with their recent moves joining Grammarly and reshaping a new kind of AI productivity suite I’m extremely curious to see how they approach the connectors topic. 

I'll be talking about progressive tool calling in Super, basically what we've been figuring out at Slite and Super while building agentic RAG for company knowledge.

I'd love to see more folks who are building in this space there.
Register here: https://luma.com/olsyxsvn

Company context is the last unsolved layer in the AI stack.
The model layer is mature. Frontier reasoning is a commodity.
And that's why, today, we're announcing Super's MCP

The moment you try to deploy that reasoning inside a real company, you hit a wall that has nothing to do with model capability but everything with context.
Models need your company's entire context to do useful work.

Every team that has tried to solve this pain.
Teams have tried individual tool connections with MCP and get a fragile stack that returns incomplete results, or more annoyingly so, times out.
Or assembling context by hand and created an expensive copy-paste workflow that defeats the purpose of having AI in the first place.

Today we're launching the accurate and exhaustive company search agent you can bring in all your AI tools to truly make them work.

Meet Super MCP.
It gives your AI a dedicated context agent that knows everything your company knows across Slack, Drive, Linear, Notion, Jira, GitHub, and twenty other tools.
→ It answers LLM's each question with real company data
→ It doesn't time out on MCP calls
→ It keeps your context clean
→ It works with Claude, ChatGPT, and Cursor out of the box.

We benchmarked it against real company queries before launching.
Claude with individual MCPs gives 24% perfect answers
Claude with the Super MCP gives 80%+ perfect answers while being 2.6x faster.

Same model. The only variable was the retrieval layer underneath.
Super MCP is live today for all Super users.

Read more about it here: https://www.super.work/mcp

And if you're trying to figure out the context layer for your team, happy to help.

Your knowledge base shouldn't be a UI. It should be infrastructure.
Every company is building AI workflows now. Sales teams use Claude to prep for calls. Engineers use it for architecture decisions. Growth teams brainstorm experiments in it. The work is real.

But the AI has nowhere to read verified team context from. And nowhere to write it back to.

Your knowledge base should be that layer. Not a wiki people visit when they're lost. The substrate your tools, your agents, your workflows operate on top of.

We just shipped Slite's MCP server. It turns your knowledge base into infrastructure that Claude, ChatGPT, and whatever comes next can read from and write to directly.

What that looks like in practice:

1/ Claude is helping you prep a board update.
Instead of you feeding it context manually, it pulls your latest strategy memo, your product roadmap, your hiring plan from Slite. It reads from your knowledge base the way a new hire would read your wiki on day one, except it does it in seconds.

2/ You finish a pricing analysis in Claude.
The conclusion, the reasoning, the trade-offs. Instead of it disappearing when you close the tab, Claude writes it back to Slite as a structured decision doc. Tagged, organized, findable by anyone on your team.

3/ You have a strategy meeting.
You're pulling insights from all your MCPs in Claude, make a plan, add your thoughts, tell Claude to make it into a detailed strategy doc, and find it already exported to Slite with the perfect formatting.

Team context gets a verified home. Tribal knowledge gets harnessed. And none of it requires anyone to open a UI.

In 2022 we published our compensation formula publicly.
We had 32 people, 25 cities, 9 countries.
4 years later, we removed a big part. Here's what we learned.

1/ Cost of life as the main criteria for our benchmark was idealistic and wrong
Our original model used a cost-of-living as the main index, which essentially answered the question "what does rent cost in Lisbon?".

While it can feel logical to aim for compensation to match the same level of life on multiple locations, it definitely does not feel fair to all, and brought hundreds of issues when negotiating salaries for existing and new teammates.

We switched from looking at cost of living to simply basing our numbers on a benchmark of actual salary data from over 1,000 venture-backed SaaS companies in our locations.

The question became "what do top companies pay a senior engineer in Portugal?" and setting a competitive number from there. Completely different framing.

2/ We kept on simplifying the structure and making it more transparent.
A comp model only works if everyone using it can explain it in 60 seconds. At first, ours couldn't pass that test.

The best way for simplicity is by exposing not only the formula but also the results. We ended up giving our team team a public datalooker sheet with all the levels/countries/roles and associated comp accessible in a few clicks.

3/ We raised our benchmark over and over as we grew, got profitable, kept on growing.
A defining moment came in 2024 when Kamsa data showed that salaries had dropped across the board.

We chose to keep our 2022 benchmarks anyway. Market data should be a guide for doing right by our people, not an excuse to quietly cut costs when the numbers give us permission to.

The rest of our calculator remained, and most of our principles still apply (go check the original piece!)

I'm sharing this because comp deserves more thoughtful attention in startups than it usually gets.

Young founders either overpay early and create expectations they can't sustain, or underpay and lose the people who built the foundation.

Both paths lead to hostile team dynamics that are incredibly hard to undo once they set in.

True agentic search in action ✨

We just shipped our latest drop on Super.work and this screen is my highlight.

It embodies our future:

One portable hub of company intelligence, accessible from all your tools. Slack, Claude, Chatgpt, Gamma, Lovable, wherever your team works and need accurate team knowledge. 

In all these, they will get
→ Deep agentic search
→ from ALL their tools
→ with the power of semantic search
→ brought at the speed of thought
→ and mixing qualitative and quantitative insights (Bigquery and snowflakes are rolling out in beta 🤫)

Check out our full changelog → super.work/changelog

Youtube can teach you to implement RAG over a weekend.
So why'd we spend 2 years?

RAG is the to-do app for the AI era. 
Every developer will build one as a pet project to learn about LLM implementation. 

There's plenty of free Youtube tutorials, wire up an open source vector store, and get answers.

But that's level 1, it only works for predictable questions.

The moment your team has real questions, and knowledge scattered across Intercom, Notion, Slack, your CRM, and 6 other tools - a single retrieval method breaks down fast.

What we've spent 2 years building at Super is an company search agents, orchestrating dozens of tools and sources, adapting to what it finds and able to search in multiple different ways.

That means:
→ An LLM query planner that decomposes your question, prepare sa first series of steps, selects filters before a single retrieval even fires
→ Multiple retrieval engines running in parallel or in sequence — semantic, keyword, SQL, MCP, and hybrid
→ Permission filtering at runtime so people only see what they're allowed to see
→ Ranking, reranking, and hallucination prevention baked in — not bolted on
→ Every answer cited, verified, and source-linked

Could you build this on n8n? No. The gap between a RAG pipeline that returns chunks and a retrieval agent that thinks about your question, searches intelligently across every tool your team uses, and verifies its own answer - that gap is the product.

We didn't build a wrapper around embeddings. We built the most thoughtful multi-tool retrieval agent we could. And we keep on adding to it.

Your customers can feel the gap between your landing page and your support calls.
That gap is your actual brand.

I don't think SaaS has reached a consensus on what brand actually means.
Everyone uses the word. Very few people mean the same thing by it. Is it your visual identity? Your positioning? Your vibe? The feeling someone gets after using your product for a month?

After 10 years of building companies, here's the working theory I've landed on. 

Brand is the extrapolation of how you actually treat your team, your customers, and what you care about when nobody's evaluating you.

1/ A good brand is downstream of your internal culture and values
At Super and Slite, I've noticed the moments where our brand felt strongest where our tone, our design, our content just happened to match how we actually behave internally.
When your team is warm, playful, and obsessively helpful behind the scenes, and that same energy shows up in your onboarding, your docs, your error messages = something clicks.
People trust it because it's consistent in a way that's hard to fake.

2/ A good brand can't be retrofitted
Where I've seen it go wrong, including in my own work, is when you start from the market. You look at competitors, pick a lane, then backfill a personality.
The illustrations look friendly. The copy sounds human.
But then a customer hops on a call and the energy is completely different.
That gap is hard to recover from.

3/ How I keep operating
Name what's actually true about your company first. How do people talk to each other? What do you default to when things get stressful? Then take that vibe and push it into every touchpoint.

It sounds obvious written down. In practice, almost nobody does it this way because it requires admitting what you actually are before deciding what you want to be.

Still figuring this out honestly. But 10 years in, this framing hasn't let me down yet.

Our median time to close bugs dropped 60% this year.
The obvious assumption is that AI coding tools did it. 
But the actual reason is a bit more subtle, and less exciting than that.

We spent 6 months last year deleting, cleaning and refactoring.
We cut 3 features that had been creating edge cases for years.
We converted everything to TypeScript, and migrated off deprecated BrowserView entirely.
We moved from ESLint to Biome and got our linting time for the entire codebase down from what was effectively infinity to 7 seconds.
We reshaped entire core parts of our codebase.

Just cleanup. That work is what made AI truly effective afterward, and I feel every day that a mature codebase with this level of quality is the unfair advantage startups will die for in the future.

When your codebase is simple, bugs are easier to find, easier to reproduce, and easier to fix, by human and agents alike.

AI coding tools amplify that simplicity because they're working with a clean surface instead of fighting through layers of dead code and undocumented edge cases.

The compounding effect showed up faster than we expected. One of our engineers spent three weeks refactoring Super's core engine with zero visible product changes. After that architectural work, the team added Bigquery as a data source in 4 days, Salesforce in 3, and the next two sources took 2 days each. That refactor felt slow in March. By November it had saved us 8 weeks of work.

PRs merged per contributor went from around 30 per quarter to 70 last year, more than 2x output with the same team and headcount.

The reverse is also worth thinking about. If your foundation is messy, AI tools help you create bugs faster and ship technical debt at a pace that would have been impossible manually. Speed is a multiplier, and it multiplies whatever is already underneath.

Most teams I talk to right now are shipping 2-3x faster with AI coding tools, and that's genuinely exciting.

But I think the conversation about what to do before you go fast is getting skipped entirely. The 6 months we spent cleaning up felt unproductive at the time, but that unglamorous work is what actually moved our metrics.

taplio

Christophe Pasquier's Linkedin Analytics