AI Agents Need Onboarding. Nobody Knows How to Do That Yet.

01 /The Concept

Pensieve built a platform for automatic AI agent onboarding – so each new agent doesn't need to be manually "briefed" on how the company works every time it's deployed.

Put differently: every other infrastructure solution lets AI agents connect to raw data. Pensieve connects AI agents to an *understanding* of how the company operates.

In that sense, Pensieve is distinct even from today's popular shared memory platforms like Mem0 – which, incidentally, has already raised $24.4 million in funding.

Those platforms let one AI agent create notes and deposit them into a shared memory layer that other agents can read. But such notes amount to a scattered collection of facts – better than nothing, but still not a coherent picture of how the business functions.

Pensieve's architecture operates across three layers:

- The bottom layer is raw data: documents, messages, and other information.

- The middle layer is a structured knowledge graph extracted from the raw data: people, projects, concepts, operating rules, and the relationships between them.

- The top layer is a live snapshot of what's currently happening at the company, overlaid on its organizational structure and active projects.

AI agents connect not to the raw data layer at the bottom – but to the top layer, which represents a live *understanding* of how the company is functioning right now.

The raw data layer updates continuously, which cascades upward through the layers – outdated information is removed, new information is added, and changed information is updated. This means agents don't operate on static onboarding knowledge that may have become stale – they work with current information that refreshes automatically.

The Pensieve platform is launching this month, and the startup is already collecting a waitlist for early customers – news that surfaced through a recent Product Hunt announcement.

Interestingly, it's not yet entirely clear how Pensieve plans to monetize. The platform itself is announced as free, with customers paying only for the AI model usage they connect through their own accounts or API keys.

Most likely, Pensieve will pursue a freemium model – offering paid premium capabilities on top of the core free platform. Those premium features could include purpose-built AI agents for common business tasks, not just additional platform capabilities.

02 /Why It Matters

Let's start with the problem today's startup set out to solve – because it's not just a new problem. It's one where it's genuinely unclear who's supposed to own it in the first place.

The Pensieve co-founder describes his team's experience with Claude Code: to get an AI coding assistant to work correctly and efficiently with a codebase, it's not enough to point it at a repository. You need a well-crafted CLAUDE.md file – one that explains what the software does, how the architecture is structured, and how that architecture maps to the files in the repo.

Without that, the AI ends up analyzing the entire codebase from scratch every session – and making its own assumptions about how things work. Assumptions that can be wrong, leading to code that misses the mark.

The same dynamic plays out when a business starts deploying AI agents into its operations. Yes, agents can analyze raw data from corporate platforms and data warehouses they're connected to – but that's like handing a new developer access to a grep search utility and calling that the end of onboarding, trusting them to start improving the codebase from there.

In practice, a new developer needs to understand what the code does, how it's organized, what lives in which files, and so on. That explanation has to happen every time someone new joins.

The same is true every time a new AI agent joins a company's operations.

Each agent needs to be onboarded – given explanations of how the company works, what it does, where the agent should pull information from, what to do with it, where to put outputs, who to escalate to when something goes wrong, and everything else needed to function properly.

Skip that onboarding and the deployment is likely to fail – the agent won't figure it out on its own. But doing that onboarding from scratch for every new agent is a genuine hassle.

And it's also an orphaned task – unclear who should own it. People who understand business operations rarely understand the technology. People who understand the technology rarely understand the business operations.

03 /Opportunities

The broader trend: AI agents need to get smarter. Not in the sense of more capable base models – that's OpenAI's, Google's, Anthropic's job. Smarter in the sense of getting up to speed faster on the specific contexts and tasks they're being deployed for.

That's a separate problem – one that base model developers can't solve, because it can only be addressed at the point of deployment. Which opens up an entirely new layer of opportunity for applied developers sitting between the model layer and the end client.

This is actually encouraging – because for a while it looked like there might not be room for anyone on the AI market except the model developers themselves.

As it turns out, the implementation layer hasn't gone anywhere – it just has to build different types of platforms and tools. Things like AI agent onboarding platforms, which is exactly what Pensieve is building.

One other startup worth mentioning in this context: Edra ([covered previously](/review/buterbrod-s-bolshimi-dengami)), which is digging in exactly the same direction. Their pitch: "Our platform learns how your business works. Then automates it." Edra surfaced a few weeks ago – but already with $30 million in funding from Sequoia Capital and a couple of other major funds, plus notable early clients including HubSpot and ASOS.

What kind of platform for deploying AI agents in enterprises could you build?