This week I listened to LatePost Chat’s episode “Agents Are Opportunities—and So Are Agent-Building Tools | Starting From OpenAI DevDay.” Henry, the guest, mapped out six major inflection points in the evolution of agent tooling: from ChatGPT’s launch at the end of 2022 to the recent wave of “computer use.” Every leap in model capability exposes a new “body part” that needs strengthening, and each gap spawns a new tool.
Listening to him felt like connecting the dots. The past two years have played out exactly that way—every few months a new capability drops, and a new crop of companies rides the wave. LangChain grabbed the earliest orchestration need, LiveKit caught the voice-interaction surge, E2B and Daytona went after secure code-execution sandboxes, and Browserbase is now surfing the computer-use moment.
That framework reminded me of the founders I met during San Francisco Tech Week. Back then I didn’t fully grasp the sequence. Looking at it now, it’s clear each of them seized their moment right as their particular node lit up.
Lessons From LangChain
Henry pointed out that LangChain evolved from a basic orchestration framework into a unicorn by doubling down on Observability & Evaluation, now pulling in tens of millions of dollars in ARR. I used to shrug at LangChain, feeling it introduced needless abstraction. Many in the community were critical too. Teams with strong in-house engineering often don’t need that layer—it can complicate integration between agents and legacy systems.
But in hindsight, that disconnect just shows how market demand and technical ideals rarely line up. The “technical” crowd I identify with is only a slice of the market. There’s a huge group of people with real use cases but limited engineering capacity. For them, whipping up an agent with a few lines of LangChain is a brilliant on-ramp.
More importantly, LangChain never stopped at the first product. They stayed glued to the market and kept shipping: LangGraph, LangSmith, and now a focus that actually sticks. That “cling to the category, pivot boldly, expand horizontally” playbook lines up perfectly with Reynold Xin’s argument from last week’s post.
The Jolt From Composio
The podcast also highlighted Composio, a tool-calling startup that raised $29 million in July. I met their CEO on a Tech Week panel. According to Henry, the whole team was based in India, tiny in headcount, and only recently relocated to San Francisco to be closer to customers.
On stage, his English was rough—I honestly struggled to catch everything. I had no idea they had just closed that round or moved countries. If I had only read the funding headline, I’d instinctively look up in awe. Meeting founders in person collapses that mystique, and that realism—neither humble nor arrogant—matters.
The real shock, though, is that Composio’s CEO blew up by relentlessly posting scrappy videos on Twitter. Even deep-infrastructure teams here host build days and hackathons to get users’ hands on their tools.
Some of my friends tried building an MCP-server index similar to mcp.so, hoping to differentiate via test reports and capture SEO traffic. It flopped. Truly useful MCP servers—those shipped by major platforms—don’t need third-party testing, and the long-tail ones aren’t worth testing. The imagined differentiation had no value, and the SEO strategy was shaky from the start.
Where Are the Users?
Put the two approaches side by side and the contrast is obvious. Among developer friends—especially AI early adopters—Google usage is way down, while the communities (Reddit, Twitter, scattered Discord servers) are booming. Instead of waiting for search traffic, it’s better to show up in the densest developer communities and pitch directly.
Search and community traffic differ most in intent purity. Someone searching “MCP” and clicking your site isn’t necessarily a real MCP user. But run a build day that requires your product, and every developer in the room is deeply hands-on, giving gold-standard feedback. In early product phases, finding core users is everything.
Here’s the painful detail: even though the team above was building an MCP index, not a single teammate was a heavy MCP user. That alone was a red flag—we barely understood real users. You either are the user or stand inches away from them. That’s Product 101, AI or not.
Meanwhile Composio’s CEO, with imperfect English, just kept showing up on Twitter—“imperfect but consistent output.” That’s the classic internet playbook: launch early, iterate fast. If a product never lands in front of its true users, it gets stuck in limbo—too afraid to ship, too invested to quit. Being criticized like LangChain means someone cares. Radio silence is the true danger.
Efficiency Isn’t Direction
This loops back to what I wrote the past two weeks. On 10.11 I mentioned Claude Code helping me crank out 5,700 lines in one night. On 10.19 I said AI let us finish three weeks of work in one. The speed boost is real and intoxicating.
But there’s a long, hard road between higher efficiency and product success. If your direction is off, better tools just help you sprint the wrong way faster. When you don’t even know where your users are, “working faster” is just frenzied motion.
On 10.05 I complained about whacking random moles and ending up with nothing. Looking back, effort wasn’t the problem, nor tool quality. We simply hadn’t answered the base questions: Who’s the user? Where are they? What do they need?
Bottom line: building great products still hinges on fundamentals—and those fundamentals matter more than ever.