Seventy-two hours until Moscone West.
TechCrunch Disrupt 2026 runs October 13–15 in San Francisco, with 200 companies in the Startup Battlefield 200, 10,000 founders and investors in the halls, and 250 speakers spread across six stages. Discounted online pricing closes when prices increase on October 13 at 8 a.m. PT. Those are the logistics. What interests me is a harder problem: how do you evaluate 200 pitch decks when a large share of them will describe something called an agent?
I spend my working life reading agent architectures, and I have developed a slightly unfair habit. When someone shows me a demo, I stop watching the output and start asking where the state lives. That single question separates the companies that have built a system from the companies that have built a prompt with good taste.
The demo is not the architecture
Agent demos are unusually easy to fake, and I mean that without accusation. A well-chosen task, a clean input, and a model with a long context window will produce something that looks like reasoning. The demo runs once, in front of an audience, on a happy path the founder rehearsed forty times. Nothing about it tells you what happens on attempt 300 with malformed input and a tool that returns a 503.
So at an event like Disrupt, where the Battlefield 200 pitches compress a company into a few minutes, the signal is not in the output. It is in the questions a founder can answer without hesitating. A few I would ask:
- What does your agent do when a tool call fails twice? Retry, escalate, or hallucinate a result and keep going?
- Where does memory live, and who decides what gets written to it? A vector store is a storage decision, not a memory architecture.
- How do you know a change improved anything? If the answer is vibes, that is an honest answer, and it tells you the stage the company is at.
- What is the unit of work your orchestrator schedules? Teams that cannot name it usually have one giant loop and a lot of hope.
Founders who have run their system in production answer these instantly, often with a war story attached. Founders who have not will reach for the model name. That reach is the tell.
Narrow beats general, and the market is learning it
One company in this year’s Battlefield 200 is dodocs.ai, which sits in document processing. I have no inside knowledge of their stack, so I will not pretend otherwise. But the category itself is instructive, because document extraction is exactly the kind of problem where agent architecture earns its keep or quietly fails.
Documents are adversarial in a boring way. Layouts shift, scans are crooked, tables span pages, and the same field appears under four different labels across three vendors. A single model pass gets you to a demo. Getting to production means validation loops, confidence thresholds, selective human review, and a feedback path that turns corrections into something the system actually uses. That is systems engineering, and it is not glamorous enough to put on a stage.
My prediction for what you will see across the 200: the companies with durable moats are narrow. They picked a domain where correctness is checkable, built an evaluation use before building features, and treat the model as a replaceable component rather than the product. The companies with thin moats will describe a general-purpose agent platform in a space where three well-funded labs are shipping the same thing for free.
Why the hallway matters more than the keynote
Six stages and 250 speakers produce a lot of polished narrative. The roundtables and breakout sessions are where the useful information leaks, because the format is small enough that someone will admit what broke. I have learned more about agent reliability from a founder describing a two-week outage than from any keynote about the future of work.
If you are attending, I would optimize for density of technical conversation over stage time. Pick ten Battlefield companies whose problem domain you understand well enough to ask a second question. The second question is where the architecture shows up. The first question gets you the pitch; the second gets you the truth.
The framing around Disrupt is about meeting startups before they go mainstream, and that is fair. But the companies that reach mainstream from this cohort will not be the ones with the best demo this week. They will be the ones whose systems degrade gracefully when the model has a bad day, because in agent products the failure mode is the product. Everything else is a very expensive autocomplete with ambition.
🕒 Published: