Hi friends! If you’re receiving this, it’s because we've either: worked together, co-invested together, or connected over the past few years with the intention of a potential partnership.
We consider you a friend of Essence and would love to keep you looped in on how we interpret what’s happening in the industry and across our portfolio.
However, if you’re not interested in receiving these emails, no worries - feel free to hit unsubscribe below.
With that out of the way, we have a lot to fill you in on!
Technical Insights
For most of software history, judging whether a product was good meant asking a human. Does it load fast, is the interface intuitive, can someone reach hello-world in five minutes? Developer Experience became a discipline because those answers determined adoption.
The same question is now being asked about a user who, at least for now, can't answer it. Agents already make up a meaningful share of the traffic hitting developer tools, APIs, and SaaS platforms. They read docs, call APIs, provision infrastructure, and decide which tool to reach for next. In their current state they don't tolerate ambiguity, don't ask for help, and don't file support tickets. They move on or they fail, and the human who deployed them often can't tell you which happened or why.
This is the increasingly pressing topic of Agent Experience (AX), a buzzword that has been building since the agent boom. Some of the current limitations in AX are real properties of software talking to software. Many are just a symptom of being in the early innings of AX.
An agent that files a structured bug report, or pauses to ask a clarifying question instead of guessing, or backs off and retries a different path when the first one stalls, is not hard to imagine. Some coding agents already do versions of this. As those behaviors mature, the bar for good AX will move. The opportunity is that the products that meet agents where they're headed get to define what the category expects.
Underneath all of it is a more existential shift in who the developer even is. You live in Claude Code or a terminal like Warp and the agent is what ventures out, provisioning infrastructure, wiring up MCPs, authenticating to integrations, and touching a dozen tools you may never open yourself. Which is great, until something breaks. Then you're debugging a system you didn't assemble, across surfaces you've never seen, with no mental map of how the agent got there. Good AX is about whether the human managing it can figure out what happened when it doesn't work as they understood.
Given that, the companies that take AX seriously are building for the agent from the first commit.
Insforge is a backend built this way. Every operation a human would perform through a dashboard, an agent performs through the CLI with structured JSON output: database provisioning, auth configuration, storage, edge functions, deployments, etc. Installing the CLI automatically deploys Agent Skills, structured context files that load directly into the agent's working memory.
Magic Patterns applies the same logic to the design-to-code handoff, historically the lossiest step in front-end work. A coding agent in Claude Code or Cursor authenticates to their remote MCP server over OAuth and pulls a design straight into the codebase. They also ship skills alongside their MCP, encoding workflows such as generating several distinct design directions off a baseline of your current UI.
If that's the net-new build side, someone has to measure AX for the existing systems. 27 points a coding agent at a developer tool with a single prompt and then measures setup time, tool calls, human interruptions, error recovery, and docs discoverability. The variance in the results is not subtle; Neon and E2B get an agent to a working hello-world in under two minutes with no human intervention, while others take ten minutes, throw cryptic errors, and require multiple rescues before limping across the line. What separates the best from the rest is clean API surfaces, machine-readable docs, and whether an error message carries enough context for an agent to self-correct rather than stall. Their bet is that as the ecosystem matures, agents will simply route around products that score poorly.
There's a hidden cost to all of this though, and both the measuring and building sides need to see it.
27 scores every tool in its benchmark using Claude. That's a logical choice since Claude Code took 46% of the most-loved vote in the Pragmatic Engineer's February survey, against 19% for Cursor and 9% for GitHub Copilot, and it leads on exactly the complex, multi-step workflows where reasoning quality decides whether an agent finishes or stalls. But when one model is the de facto runtime for the agents judging your product, your AX score is a Claude score.
The build side has the same gravity. Agent Skills is an open standard Anthropic published, already adopted by 40+ development tools. MCP also started at Anthropic, but it's since gone cross-client since OpenAI and Google both support it. Skills are earlier and more concentrated.
So a category where one model's preferences decide which tools get picked is fragile the moment that model's behavior shifts. Products that score well get recommended by agents, then adopted by the humans who trust those recommendations. Ones that don't adapt to the rubric get skipped, which is a harder failure to see and an even harder one to reverse.
The teams we're watching in this space are building for a user that reads everything, forgets nothing, and rarely admits confusion. Win that user and the humans come with them, because the human is increasingly just asking their agent what to use.
Portfolio Updates
CodeRabbit has raised $143M at a $1.5B valuation. They’re betting that the rapid adoption of AI for development will increase demand for software monitoring and code fixes as companies have to audit AI-generated software. Read more here.
Ollama is rolling out support this week to host Meta’s Muse Glimmer (30B open-weight model), which has been designed to run autonomous AI agents directly on consumer hardware, pushing agentic workloads that normally depend on cloud infrastructure onto Macs and PCs. Read more here.
MotherDuck announced "Guides," a new context layer for AI agents working with data warehouses, claiming a 10x reduction in AI-assisted analytics costs by enabling distributed, organized context that prevents redundant token consumption across multiple agents. Read more here.
DevTools Worth Trying
Cline has made NVIDIA’s Nemotron 3.5 Lightning, a 30B MOE model with 3B active parameters and 1M token context window, free to use! The model is designed for always-on agents that need to move through many model calls quickly, making it a strong fit for Cline's agentic coding loop.
Warp just released their Agent CLI, a coding agent for engineers who live in the terminal. It’s a state of the art, terminal-native coding agent CLI that keeps you in flow across sessions, remote systems, debuggers, and REPLs. Now available in any terminal. More on the release here.
OpenInfer just launched their Cloud, a hosted multi-tenant AI token platform built on their dynamic inference operating system. It hosts popular open-source AI models and delivers lower costs per token without requiring code rewrites or infrastructure setup. Read more here.
Podcast Episodes
Here are our most recent podcast episodes from The Infra Pod featuring William from Buoyant and the Open Source Startups Podcast featuring Dotta (Nate) from Paperclip:
Open Jobs
Insforge is hiring top-tier talent, and is willing to pay $$$ for a referral:

See more details here.
Many of our other portfolio companies are hiring across similar roles as well, explore them all on our Essence Careers page.
Events
MotherDuck is hosting a live virtual demo on August 18th on how to build the entire stack around your AI-generated SQL queries: schemas, dbt models, metric definitions, etc. The entire system will be built in under an hour and a live ingestion pipeline for a real REST API will be produced to show how easy it is to keep AI agents honest. Check it out here.
Ollama is cohosting an evening meetup in Dogpatch with OpenAI and vLLM on August 27th to discuss gpt-oss. Topics covered include what it is designed for, expected usage, implementation, and how Ollama enables its capabilities. Check it out here.
LanceDB is hosting a one-day summit in SF on November 5th about the data systems beneath the most cutting edge AI models. Meet the people running the most ambitious training runs and hear about their architectures, tradeoffs, failures, and lessons learned. Check it out here.
Thanks for reading! We'll be back next month with more updates.
— The Essence Team

Asim Moinuddin, Tim Chen, Naomi Walker-Garrett


