Author: @danbeksha

Translation: Peggy, BlockBeats

Editor's note: AI is entering enterprises, but the real question is not "should we use agents," but whether these agents can understand the company itself.

This article discusses a deeper issue through the lens of the author's first 100 days at Ramp: a fast-paced company cannot rely solely on new employees to read documents, ask colleagues, and fill in the context slowly, nor can each AI tool operate independently. What truly matters is building a continuously updated "company brain" that consolidates meetings, documents, Slack discussions, customer feedback, and product decisions, allowing both new hires and agents to start from the same contextual knowledge.

When context is systematized, onboarding is no longer just a long adjustment period, and AI is no longer a set of isolated tools. The value of enterprise AI may ultimately lie not in how many agents are deployed, but in whether a company can first establish a trustworthy, readable, and reusable knowledge base.

The following is the original text:

In a 4×100 meter relay race, the outcome is often determined not by the entire duration but compressed into a 20-meter exchange zone. Runners must complete the baton pass at high speed: if the incoming runner starts too early, the baton can drop; if they start too late, the outgoing runner has to slow down, and the entire team can lose its advantage in an instant. If the exchange itself is not precise—any error in hand position, angle, or timing can lead to dropping the baton.

A team can have the fastest runner on the field but still lose in those 20 meters. Speed is important, and the exchange is important. What truly determines victory is whether both can coexist simultaneously.

Every job handover I have witnessed has been, in essence, a relay race, but one of the runners is still at the starting block. A new hire starts on Monday with everything at zero; the organization, however, does not slow down but continues to function at its original pace. Thus, the new hire must rely on reading documents, lurking in Slack, repeatedly asking the same questions, and spending three months grasping the organization's operational model before they finally become "useful."

We often view this gap as a matter of time, as if, given enough time, a new hire will naturally catch up. But that is not the case. The gap must either be addressed by a system, or it will persist indefinitely.

Context is the true handover system of an organization

I have been at Ramp for about 100 days. Before this, I worked at Plaid for five years, familiar with every product, every customer story, and the background behind every decision. I could recount these stories without hesitation. However, upon arriving at Ramp, I knew almost nothing about any of this.

Yet the core of product marketing is storytelling. If you don't know the characters, plots, and cause-and-effect relationships in the stories, it is impossible to tell the story well.

From day one, my goal has been to build an AI-native product marketing organization. However, to achieve this without context, I first needed to expand my knowledge base—that is, the "context layer" that supports all work.

Ramp is a company known for its speed. There is no room for "catching up slowly next quarter." The company releases, iterates, and pushes forward every week. You either keep up with the pace, or you become an added cost in the organization's operations.

At the same time, I am undergoing another layer of onboarding. Ramp is fast, but AI changes even faster, and I must learn both a new company and a new work environment simultaneously. I am not an engineer; the last time I used a terminal was during a computer class in college. This means I need to catch up on organizational context while adapting to a new AI workflow, which intensifies the difficulty.

Ultimately, what liberated me from this pressure was not completing a specific article, product launch, or workflow but treating "context" itself as a deliverable. As long as the context layer is built correctly, all subsequent work will become cheaper.

Thus, I began to construct something truly scalable: a system that helps me quickly catch up like a great wiki helps researchers. By the third week, it could draft content based on my notes; by the eighth week, it could summarize meetings I did not attend. Learning and catching up did not disappear, but as the system continued to populate, their costs began to decrease day by day.

The personal version of this idea has actually existed for a while. Karpathy, a former head of AI at Tesla and one of the founding members of OpenAI, wrote an article in April describing what he called a "personal LLM knowledge base": a folder containing raw input, including papers, articles, transcripts, and personal notes; an LLM that generates a wiki based on this material; and a front-end editor like Obsidian. When the materials accumulate to about 100 articles, the LLM can answer complex questions around the personal corpus without needing complicated retrieval techniques.

His judgment is that there is an opportunity to create a truly outstanding new product here, rather than just a collection of ad-hoc scripts.

While the personal version exists today, the company version does not. That is precisely the problem.

Generally, here is a system I built during my first 100 days. It is not yet refined, but together they make up the "connective tissue" within the organization.

The core is an Obsidian vault, read and written by Claude. Meeting transcriptions, documents, public opinions, and personal notes that I have encountered all enter this knowledge base. When I ask, "What did Geoff and I decide about the homepage three weeks ago?" it looks for answers from this vault rather than relying on the model's generalized memory.

To continuously input content into this vault, Granola defaults to recording every meeting and archives transcriptions overnight. Therefore, a meeting I missed on Monday can already be queried by Wednesday. To help others in the company catch up, I choose to make my work public—most of the content I am building appears first in the #team-pmm or related release project channels before entering Notion documents. The building process itself serves as a synchronization mechanism.

Above this vault, there is a small named skill library that agents can call upon as needed. One skill can generate agendas based on my last four meetings with someone; another skill can scan product dynamics on Slack for the week and turn them into article topics. Each skill is about 200 lines of markdown, replacing tasks that previously required manual completion.

Additionally, I built a dynamic product roadmap based on Ramp's internal application platform. It pulls from the same context layer, so it does not expire as it was never a static document to begin with. There is also a morning summary sent to my Slack direct message every day at 8 AM: what launched yesterday, where things got stuck, and what requires my response. All this content is organized while I am sleeping.

Individually, these elements may not be impressive. But together, they provide a workable answer: what would a company have if it also possessed the kind of wiki Karpathy mentioned?

You can call it a wiki, a graph, a context layer, or a company brain. The name doesn't matter; functionality does. It must be capable of absorbing all signals the company has already generated: meetings, Slack discussions, documents, code, transcripts, customer calls, and key decisions, and must remain continuously updated without relying on manual maintenance. It must also be the first thing every new employee and new agent reads before they start working.

If a new employee joins tomorrow, what should they read on their first day? If the real answer is a Notion document from 2024, plus a now-defunct Confluence link, then essentially, it is making them pick up the baton from a standstill.

From single-point tools to a company brain, the true gap of AI

Today, the primary way AI enters enterprises still relies on forward-deployed engineers. Whether it is OpenAI, Anthropic, or large consulting firms, they choose to build specific workflows on top of models.

These tasks are real and valuable. But they still linger in the "chatbot era" of enterprise AI: narrow tools encapsulated around specific tasks, useful in isolation but not integrated into a system that can yield compounded benefits.

The true "company brain" has not yet emerged. Customer service agents and HR onboarding agents may have been built in different months by different teams. They do not know what was decided at the last all-hands meeting, how the company understands its market, nor what judgments the sales lead made at the last management offsite. Each agent is merely a chatbot with specific responsibilities, but they do not share the same brain.

That is the largest gap currently. Outside of laboratories, there are hardly any people building products around this issue.

If you want to build a team or start a company in 2026, the order of operations is already different from 2022. First, write context documents, then install tools. Record every meeting. Build the wiki before building the dashboard. Deliver skills, not slides. Have new employees read the wiki on their first day and start contributing to it on their second. Hire and promote those who can keep the "company brain" running, and also reuse agents that can read the company brain effectively.

Context is not a secondary project. It is the foundation that truly allows all AI investments to generate returns.

I am currently building part of this at Ramp: wikies, skill libraries, applications that pull from the same context layer, and organizational mechanisms that continuously input content. It is still small and early. If you are also trying to construct a company-level version elsewhere, I would love to exchange experiences. More useful than a reliable brain is having two brains in the same room.

Back to the relay race. The true condition for victory is neither the cleanest handover nor the fastest leg but rather both occurring simultaneously within the same 20 meters.

A new employee reads the company brain and then begins to sprint. A new agent reads the company brain and then begins to work. A new customer accesses the company brain and is in operational mode from day one.

When the term "ramp-up" no longer has meaning, we will know we have done it right.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。