The bar for good product is competitive and continuously getting higher. It's a privilege to build and have a front seat to frontier application layer AI product
Sharing my observations about what wins, and what the industry is demanding of its vendors
beautifully simple piece written by langchain CEO
it breaks down how we've all seen a convergence in what problems and tools are table stakes
- tools
- skills
- compaction
etc.
naturally, we are identifying where it's taking exponential effort to tackle within your own tech stack
- personalization
- response cadence
- information retention
- maintenance over time
all of these point to another "instinct" moment - this is fundamentally a harness / context architecture problem that can be solved by putting different techniques together and experimenting for what works
super exciting time to build
langchain.com/blog/context-e…
Not surprised, everyone indexed on FDEs being a sign of bad PMF
In reality, day to day FDE is the solution to aggressively custom enterprise builds that are expected because of how competitive the market is
Not only that, once you are embedded in enterprise customers, there is so much demand for new use case experimentation like long horizon agents and outbound that having someone embedded means you can grow ACV much faster
It’s true that using FDEs to overcome bad PMF is still suboptimal, but this brings to light what is changing about traditional software in light of lightning fast coding
Distribution and relationships are the moat
and it begins.
have been and will continue to be bullish on startups building insurance policies and support for ai vendors.
have been signaling this opportunity since over a year ago when i was investing in the space. huge to see this today.
nytimes.com/2026/11/07/busin…
navigating large amounts of context efficiently will be a competitive moat
when you break down some of the core agent workflows that need accuracy, you start to realize how much alpha there is in using frontier techniques
perplexity's deep research agent naturally requires agents to be: 1) accurate, 2) reliable, and 3) consistent
it's no surprise that about a month ago they released innovations on how to efficiently pull context based on queries
quick tldr; their novel embedding model is trained to create loss scores based on both relevance within a document and relevance over all documents. this key, pair is done per token and can then be sized and reevaluated dynamically
that is really good for use cases where you need to find correct information inside dense documents that require context from the full document (hint hint, deep research)
their investment into this architecture no doubt improves their efficiency and reduces their cost burden per query
really good stuff
perplexity.ai/hub/blog/conte…
i'm increasingly impressed with Devin from Cognition
i think the way it manages context is quite effective (though it's pretty expensive)
from reading their pieces on how they do context management, their discovery into multi-agent architectures from a few months back was pretty forward looking - the biggest takeaways:
after trial and error, they found emergent success with a "smart friend" approach, where a primary, less powerful model delegates to a more powerful model (this is the inverse of what you'd expect)
clear communication between agents remains an open research problem
they prompted the smart friend to overscope: take in more context, consider pushing back, and go beyond the primary model's initial question
the hard part was knowing each model's strengths
tightly scoped manager Devins don't work when the manager lacks context, which is interesting because it's so human
map-reduce was the best implementation
--
side note - I love when engineering blogs take accountability, this is @walden_yan updating his view from a previous post in 2025
great stuff from Cognition
No matter what side you’re on, these midterms are a referendum on the problems Americans ACTUALLY want solved
As someone building in AI, it’s a moment of reflection for me that AI has yet to reach its potential panacea for the every day person
There’s still so much to build
how Sanity built their internal agent harness for the whole company (including nontechnical users) is awesome:
- they use jev as a model router, something that i think will become popular (either jev or other tool)
- they use doc for agent instructions instead of github
- they are laser focused on making the access point (the "@") easy to use and handle the routing on the backend
- they use the idea of harness building and skill building to bring people together cross functionally. i think this is super underrated in larger companies (sanity is now 300+)
highly recommend taking a look sanity.io/engineering/how-we…
This is pretty insane for a couple of reasons:
- this brings into scope potential for mech interp at scale without requiring as much intensive effort
- it could shift the competitive dynamics in favor of stronger, larger models with larger context sizes that can make broader strategic edits
- automated reasoning is a lot more reasonable if this continues to hold