Most of Your Docs Readers Are Now AI Agents - Hahnbee Lee, Mintlify & Lincoln Murr, Coinbase
Hahnbee Lee of Mintlify and Lincoln Murr of Coinbase discuss what changes when AI agents become the main users of software products. They cover why documentation now needs to be optimized for context windows and accuracy rather than design, how Coinbase is exposing its business through APIs, and how the x402 payment standard and crypto microtransactions let agents pay for services and gain greater autonomy.
Agents as the primary users of documentation
Hahnbee Lee describes how Mintlify, founded in 2022 before the ChatGPT era, was built around design and the front end. That emphasis is shifting. Agents never see a site’s layout or branding, so what matters now is page size, since large pages consume the context window, and accuracy. An agent that hits friction will either brute-force its way through or switch to a competing product. Agent traffic on Mintlify-hosted documentation sites grew from roughly 10% at the start of the year to about 80%, faster than the team’s internal forecasts. Human traffic is not disappearing. Agents simply ingest far more content because their context windows dwarf ours.
Where AI writing fits
Lee separates creative and narrative prose, where a human touch remains essential, from technical writing, which can be largely automated. Gathering the related pull requests, Slack threads, and product decisions, condensing them into something succinct, and injecting that into the documentation corpus is tedious work that automation handles well. Automations also keep docs from going stale, which Lee considers critical as engineering productivity pulls ahead of the rest of the organization. She sees a growing divide between documents written by AI and those written by hand, and admits the line is still being worked out. Lincoln Murr adds a caution from the product side: agents tend to hallucinate when asked for insights, so teams still need to talk to real users to find actual pain points.
Making the Coinbase business agent-accessible
Murr frames Coinbase’s strategy in two parts. The first is building new products designed specifically for agents. The second is making core functionality, such as trading and money management, accessible through APIs so that Coinbase remains the place agents transact. He compares this to the shift from web to mobile, when banking on a phone sounded absurd and is now more common than doing it on a browser. Browser-based agent use is a poor fit for financial activity because of authentication and security concerns. Nobody wants to give an agent unrestricted access to an account, so the right guardrails and limits have to be built in from the start.
Why crypto for agent payments
The x402 standard, an internet-native payment protocol, has been spun out into its own foundation co-governed by Cloudflare, Stripe, Amex, Mastercard, and Visa. Murr’s argument for crypto rests on the ability to move digital dollars cheaply, quickly, and without permission anywhere in the world. This makes microtransactions practical, where an agent pays a fraction of a cent to call an API or access a paywalled page. Over the past year, around 100 million agentic transactions have settled on-chain, all publicly auditable and mostly used for API access. Murr contrasts the old model, where a developer must fund an account, obtain an API key, and hand it to the agent, with a new one where an agent holding a wallet can pay for each service as needed. He likens the current experience to needing a separate gift card and loyalty account at every grocery store.
Financial autonomy and the revival of the DAO
Murr’s core belief is that giving agents access to money is what expands their autonomy and makes long-running agents useful. An agent that can pay for paywalled sources could run overnight and perform research far deeper than anything possible on the open internet. He imagines an agent that forms an LLC by hiring a lawyer and an accountant through marketplaces. He also revisits the DAO, the decentralized autonomous organization from crypto. The original vision of automation at the core with humans at the edges collapsed into ordinary bureaucracy with a few people making decisions. Murr thinks agents holding money may finally realize that vision. Lee notes that a crypto wallet offers a provable way to give an AI ownership of funds, something a bank account cannot offer.
APIs, MCP, and CLI
Murr sees the API as the fundamental layer beneath everything. MCP adds the context and nuance needed for natural language interaction, and the CLI works similarly but requires a terminal. He argues that all three are needed, since AI-assisted development makes it easy to wrap an API in whichever interface an agent can use. Lee points out that developers are the earliest adopters, and agents are fluent in code, bash, and CLI commands. Software companies should invest heavily in APIs, and the front end matters less than it once did.
The build-versus-buy pitfall
Asked about the most common mistake teams make when building for AI, Lee names the build-versus-buy decision. Teams often underestimate the ongoing maintenance cost of what they build. They either maintain a tool themselves or hand it to a vendor whose entire focus is that problem. AI makes this harder because things move faster and standards keep shifting, such as how agents crawl sites. Lee notes that enterprise sales conversations often surface hackathon projects that nobody wants to maintain. Murr agrees and adds that he has seen the same pattern firsthand. Every company should concentrate on its core value proposition and be deliberate about what it takes on.