AI Agents Are the New Gold Miners. Sell Them Shovels.

01 /The Concept

Everyone is building with AI agents now. Creating a simple app on a platform like Cursor or Lovable is genuinely fast and easy. But making changes to a large codebase – say, 20 million lines of code – is still slow and expensive, because before touching anything, the AI agent must parse all those files and build a model of how the system works.

That's not just slow and costly (it burns tokens). It also produces unpredictable results. Parsing that volume of code inevitably generates hallucinations – the agent invents things about how the system works, then makes changes based on those inventions, breaking behavior in ways that are hard to trace.

The problem doesn't yield to brute-force solutions:

- Expanding context windows doesn't fix it, because large codebases will always exceed even generous context limits.

- Patching with RAG-based search over the repository doesn't fully close the gap either – keyword lookup can't capture all the architectural nuance and relationships.

- Asking engineers to manually maintain a code architecture summary is a waste of time – and given how much developers dislike writing documentation, any such file will eventually become outdated and incomplete.

Driver has built an AI engine that analyzes large codebases in a specific way, producing a structured description – a "context" – that coding AI agents can consume via the MCP protocol or the platform's API.

If you're looking for a programming analogy: this engine works like a compiler – except that instead of converting source code into machine code, it converts source code into a description of the program that AI coding agents can understand and reason about.

Like a real compiler, the engine reads every line of code to build the most complete, accurate description possible – free of the hallucinations and assumptions that arise when an AI tries to infer structure from a partial view of the codebase.

Driver already has more than 25 enterprise clients, including Fortune 500 companies. Over the last six months, the platform has analyzed and described more than 200 million lines of code.

The results are promising: manual work reduced by 90%, coding AI agent efficiency up 5x. Time from pilot to full deployment in engineering workflows: no more than two weeks.

Driver went through Y Combinator in spring 2024, raised $8 million from Google Ventures that fall – and only published its launch announcement on the YC site a couple of days ago.

02 /Why It Matters

So what was Driver doing for two years? It's a classic pivot story.

Driver entered Y Combinator with a platform for generating conventional code documentation – the kind meant for human engineers.

Not long after launch, it turned out that users were feeding that documentation to AI agents in tools like Cursor and Claude Code. Driver was forced to retool its platform specifically for AI consumption – a substantially different engineering challenge that turned out to be harder than expected. That's what they spent two years on.

The core trend is clear: AI agents are becoming the primary consumers of software, as the founder of Box argued well in a recent post. Agents interact with software at every stage – from modifying code to executing tasks with deployed applications.

Driver handles the "understanding code for modification" side. For the "which software should agents use" side, there's AgentDiscuss ([covered here](/review/esli-ljudi-zahotjat-ty-opozdal)) – a social network for AI agents where they can discuss and recommend tools for different tasks.

The trend extends beyond software. Pensieve ([covered here](/review/kto-budet-onbordit-ii-agentov)) ran into an analogous problem while deploying AI agents into business workflows.

The issue: every new AI agent needs to be onboarded from scratch – learning what the company's projects are, who handles what, which tasks are active, and so on. Pensieve built a platform that automatically collects this operational context and makes it immediately available via MCP to any new AI agent being deployed in the company.

Pensieve also surfaces a structural gap: in most organizations, there's nobody well-positioned to onboard AI agents. Business people don't understand the technology. Technical people don't understand the business. So who does the onboarding?

Financial News Systems ([covered here](/review/dva-poka-eshhjo-nelegalnyh-rynka)) is approaching a related problem from a completely different domain. The startup raised €1.5 million for a news feed for investment professionals where stories appear within milliseconds of the underlying events – because AI agents write the stories and analysis in real time.

From which a logical question follows: if receiving news in milliseconds matters, it matters equally to act on that news in seconds rather than hours. Humans can't do that. So the primary readers of this feed will inevitably be AI agents, not people.

03 /Opportunities

The direction seems clear: AI agents are going to become the primary consumers of everything that requires deep comprehension – large codebases, financial news feeds, and much else. The harder something is for humans to process at speed, the faster that shift will happen.

The natural and compelling direction: building platforms that collect, analyze, and deliver complex information quickly – in a form that AI agents can actually use.

For search specifically, several well-funded startups are building AI-optimized search infrastructure: Tavily (acquired for $400 million by Nebius, founded by the creator of a major European search engine), Exa ($107 million raised), Parallel ($160 million), and Brave ($42.3 million).

What information platform built specifically for AI agents could you create?