Patrick Chase

FDEs are uniquely suited to be founders. @getserval announced an Serval Start: a two-year program for FDEs who want to be founders in the future. This is a truly incredible opportunity to see enterprise AI agents up close, work with an exceptional team, and build out network of peers and investors. Excited to be involved! Apply here: serval.com/serval-start

Parallel Agents just shipped. Mix and match any agent, run them all at once, and manage everything from a new Threads Sidebar. Claude Agent, Codex, Zed's agent, or anything on ACP. One window, across all your projects. zed.dev/blog/parallel-…

@sarahcat21 @modal 💚

In the past year, the number of sandbox environments has exploded; sandboxes are now a core primitive for agent developers. @modal’s sandbox product started as a weekend prototype 3 years ago, but now handles 100Ks of concurrent environments. More on how they built it: Many people think sandboxes exist exclusively for coding agent inference. @Lovable, @tryramp, and @cursor_ai have all discussed their sandbox requirements when deploying coding and background agents in recent posts. However, a new use case is emerging that is often more infrastructure-intensive and complex: reinforcement learning. Today's RL-postrained coding agents benefit from verifiable rewards. The model generates code during rollout, runs it against a test harness, and receives objective feedback (i.e., did the tests pass?). The total number of executions scales multiplicatively with the number of tasks, sampled trajectories per task, and steps within each trajectory. What starts as a unit test quickly turns into a stress test…for the researcher. Faster sandbox provisioning increases the rate at which fresh, on-policy data can be incorporated into training, directly impacting learning efficiency. One of Modal's customers, a major AI lab, is already running on the order of 100,000 concurrent sandboxes for RL workloads, with a stated goal of reaching 1 million. @AIatMeta recently released an open-weight model for code generation that was RL-post-trained with Modal sandboxes. So why sandboxes instead of VMs? Primarily isolation — sandbox logs routinely show reward hacking by attempting destructive behavior. Additionally, the orchestration of VMs (spinning up images, capturing logs, proxying commands) is a PITA. And then there’s resource efficiency — each sandbox only needs a slice of a CPU core and a bit of RAM. The initial v1 was built in a weekend by @akshat_b (anyone who knows him won’t be surprised). But Akshat and his team had already spent years building the right primitives like fast container startup, a custom filesystem (two actually), gVisor-based isolation, and a scheduling layer that packs workloads efficiently across a fleet of machines. Most of the V1 engineering work was API design — figuring out the right SDK for an imperative, agent-driven model of compute. Scaling past the weekend prototype has been more arduous. There were a few core problems the Modal team needed to deal with: First, scheduling becomes a real-time systems problem at this scale. Modal runs across multiple regions (unlike some competitors who operate within a single region and conveniently omit this when reporting cold-start benchmarks). Then there are primitives around sandboxes. Modal invested heavily in their storage layer, including filesystem snapshots, directory snapshots, memory snapshots, and volumes, because agent workflows are inherently stateful, and these primitives make that state explicit and controllable. They also support GPU-backed sandboxes, which expand the set of possible workloads but add another layer of scheduling complexity. Modal sandboxes are quickly becoming the industry-leading product in this space. At the time of writing, Modal can spin up hundreds of sandboxes per second for a single customer. These colossal numbers are driven by real RL training workloads where sandbox throughput is a bottleneck to model improvement. Read more in my full post here: amplifypartners.com/blog-posts/beh…

"Talent is the only durable moat." @jakeserval on how a motivated competitor with access to the latest AI tools can copy everything within a year. The only thing that compounds in a way that can't be replicated is the people you build with.

This week on Unsupervised Learning, @jacobeffron and I sat down with @jakeserval, co-founder & CEO of @getserval. Serval is going directly after ServiceNow in ITSM and already working with companies like Notion, Clay, Abridge, Fox, Mercor, and Verkada. We partnered with Serval at the Series A. What’s stood out most over the last year is their speed of execution. We get into how they’re winning customers and talent so quickly, including: ▪️ Why building a system of record beats layering on top ▪️ The "mirror architecture" that lets Serval land enterprise customers ▪️ Why ITSM is more vulnerable to AI disruption than other verticals ▪️ The Future IT Stack when agents submit their own requests ▪️ The AI-native org chart ▪️ Why recruiting is the #1 job of every Serval employee ▪️ The Dream Team Draft: recruiting during hypergrowth YouTube: youtu.be/Q0bxRANHjFY Spotify: bit.ly/4m4PJRX Apple: bit.ly/3POsYp8 0:00 Intro 1:25 What is Serval? 4:51 Early Doubts and Strategy 6:34 AI Tailwinds in ITSM 8:04 Competing with ServiceNow 9:41 Why ITSM Is Vulnerable 11:52 Automation via Codegen 16:27 Critical Guardrails 28:32 Internal Support Complexity 30:24 Hiring as the Moat 31:44 Dream Team Recruiting 33:49 Managers vs Super ICs 36:44 Junior Engineers and AI Native Workflows 43:13 Quickfire

Serval is #1 on the Enterprise Tech 30. What a surreal moment. Special to see so many of our customers here like @clay, @vercel, @togethercompute, and many more we'll announce soon. I’m so grateful for the entire @getserval team that powers this rocket ship, especially my cofounder Alex McLeod and COO @TatianaBirgisso, who have assembled unrivaled teams across engineering and GTM. It’s Serval now. newcomer.co/p/mintlify-ser…

as of last week, agents are creating more cells in @_hex_tech than humans directly

💚💚💚

@sahirazam @livekit @dsa @davidzh 🔥

Software is learning to listen, speak, and act in real time. The infrastructure for that barely exists. That's why we're so excited to lead @livekit's $100M Series C. @dsa and @davidzh are building the real-time layer for the future of software and agents!

@dsa ❤️❤️

Today is a day I’ll never forget. I grew up in Cupertino. My dad was a tech founder in the 80s/90s. I was in YC S07. LiveKit is my 5th company. The first 4 didn’t work out. I’ve had a lot of advantages — it still took 20 years to get here. Founders: keep taking shots.

Huge congrats to @dsa @davidzh and the entire @livekit team on reaching unicorn status and announcing the Series C! Their voice AI platform is 🤯 and powering apps that see, hear, and speak for the leading AI companies. So grateful to be on the journey with you all!

This is a big deal. Most of the voice AI community says that speech-to-speech models aren’t ready for production use cases. Two reasons: 1. Reliability (more hallucinations) 2. Cost (s2s models are expensive) On (1), Grok’s Voice Agent API is already running at large scale across Grok’s apps, in Tesla vehicles, and in call centers. There’s more work to do, of course, but progress is being made quickly. On (2), you get SOTA performance for $0.05/minute, which meets or beats aggregate cascade model (STT+LLM+TTS) pricing. Excited to partner with @xai on this launch — you can build a custom Grok Voice Agent with workflows, tool calling, the whole shebang, in a few lines of @livekit code. The future of speech-to-speech is bright!

Huge congrats @tryqualified and @alexbard !!!

Huge congrats to @getserval on reaching unicorn status! @jakeserval, Alex McLeod, @TatianaBirgisso and team are executing at mind-blowing speed and @Redpoint is grateful to be part of the journey! In just the last few months they have: 🤖 Helped customers like Mercor, Perplexity, Clay, Verkada, and Abridge automate 50% of IT requests 📈 Doubled workflows run on serval 🚀 Grown the team 3x and hired some truly incredible folks across the board Most importantly companies are now moving their IT systems of record to Serval. So excited to work with @biad_anas and @gradypb! More about the round here ⬇️ serval.com/updates/serval…

How fast can you go from question to answer? At Hex, we've been obsessed with collapsing that timeline. But in a conversation with @sarahdingwang from @a16z, our CEO, @barrald, talks about a future with more ambitious goals: What if time to insight could go 𝘯𝘦𝘨𝘢𝘵𝘪𝘷𝘦? Before you even ask to think about something, the update comes to you. We call it ambient analytics — a future where dashboards and data tools anticipate what you need and surface it before you even ask.

Will we eventually stop looking at source code? Is the death of the IDE imminent? Not if @nathansobo of @zeddotdev has his way. Nathan has been on a lifelong quest to build the perfect tool for his craft, first creating Atom at @github, and now founding Zed. On today's episode of Training Data, Nathan shares his conviction that source code is still the clearest language humans have for expressing complex ideas, AI or no AI, and a built-for-purpose GUI will always be the best way to inspect, navigate, write, and otherwise interact with that code. Listen to the full episode to hear his view on why the IDE GUI matters more than ever for AI, and where human + agent collaboration is headed.