Hi friends! Welcome back to another installment of our newsletter, we have a lot to fill you in on.

Technical Insights

The center of gravity in AI infrastructure has shifted back toward where agents actually meet developers: on their machines and inside their existing toolchains.

When agents write more of the code, the terminal, editor, debugger, and local data stop being incidental developer conveniences. They become part of the agent architecture.

Warp and Cline are two of the clearest signals, and they're carving out this territory alongside Claude Code, Cursor, and Codex. SpaceX’s agreement to acquire Cursor for $60 billion is a good indicator of how much is at stake in owning this interface. Warp is turning the terminal into an agentic development environment where developers can run and coordinate coding agents from the interface they already use to operate their systems. Cline lives inside the editor and terminal, where it can inspect files, modify code, execute commands, use external tools through MCP, and ask for approval before taking action. These products are not simply putting chat boxes next to code. They are turning the developer machine into the cockpit from which agents understand and manipulate the software environment.

The data layer is moving in the same direction.

Turso is rebuilding the application database around local reads and writes with explicit synchronization to the cloud. Rather than treating the local database as a disposable cache in front of a remote source of truth, Turso Sync allows state to originate locally and move outward through explicit push/pull operations. That model is especially compelling for agents, which continuously generate intermediate state and need fast access to working context.

LanceDB is doing something similar for retrieval, multimodal data, and agent memory. Developers can start with an embedded database running in-process, then move to distributed infrastructure when the workload requires it. Its branching model adds isolated, writable lines of table history, allowing a developer or agent to experiment with new data, indexes, schemas, or reprocessing steps without disturbing main. It is not full Git-style merging yet, but it brings the familiar idea of branching into a category where models are constantly generating and modifying state.

Modal clarifies the boundary of the local-first thesis. A developer can initiate a workload from a local Python program and debug it through a workflow that feels local, while the actual function, sandbox, batch job, or GPU process executes remotely on Modal’s infrastructure. The laptop becomes the control point without becoming the compute boundary.

That is the architecture beginning to emerge:

  1. A local control-and-state layer. The editor, terminal, model runtime, operational database, and memory system increasingly live on or immediately beside the developer machine. This is where the agent receives context, takes instructions, changes files, stores working state, and requests permission to act.

  2. An elastic execution layer. Expensive inference, large batch jobs, secure sandboxes, and production workloads still run in the cloud, but they are invoked from a local development loop rather than forcing the developer to relocate the entire workflow into a remote environment.

Ollama’s recently announced $65 million Series B is a clear proof point. Founded by the same team that helped build Docker Desktop, Ollama has grown to ~9M users by doing for local AI models roughly what Docker did for containers a decade earlier. That does not mean cloud models are going away. Ollama itself now spans offline and hosted usage. It does mean the model runtime on the developer’s machine has become valuable infrastructure rather than a hobbyist edge case.

Company Updates

DeepTrust was named a "Pioneer" in Gartner's Emerging Market Quadrant for Deepfake Detection — Startup Vendors report, recognized for going beyond detection to address deepfake threats in a differentiated way. Read the report here.

Encore has raised €5.9M, more than doubling its valuation to €37.2M. Founded by former Spotify engineers Marcus Kohlberg and André Eriksson, the company lets teams and their AI agents define backend infrastructure directly in code instead of wrestling with Terraform and YAML, and now counts 200+ engineering teams shipping on it, including Groupon, who cite a 2-3x increase in development speed. Read more here.

Sardine has announced a new integration with Modern Treasury, a payments operations platform, bringing real-time fraud monitoring and transaction risk infrastructure natively into Modern Treasury's platform. The partnership covers ACH, Wire, RTP, FedNow, Push-to-Card, and stablecoin payment rails. Read more here.

DevTools Worth Trying

Restate shipped two big releases last week. BYOC (Bring Your Own Cloud) deploys fully managed Restate directly into your own VPC, private, secure, and up to 10x cheaper than traditional durable execution at high volume. And Restate 1.7 ships Virtual Queues, a new flow-control primitive that lets you set hierarchical concurrency limits (org, then team, then user) across agents and services in real time, dynamically adjustable via one API call. Watch co-founder Stephan Ewen introduce BYOC or read more on 1.7 here.

OneWill just launched Agent Retcons, giving coding agents the time-traveling abilities of databases: autosave, rewind, retcon, and handoff around any agent (Claude, Codex, agy) with no source code or agent changes, so every irreversible action is either reversible or gets your sign-off. Get early access!

Cline added day-one support for Z.ai's GLM-5.2, the first open-weights model to break 80% on Terminal-Bench, meaning frontier-level coding performance without the frontier price tag or the closed-weights lock-in. If you've been coding with GPT or Claude models and wondering what the open-weights world has caught up to, this is the moment to find out.

Podcast Episodes

Here are our most recent podcast episodes from The Infra Pod with Tyler Wells, CTO of BrainGrid, and the Open Source Startups Podcast with Pavlo Baron of Platform Engineering Labs, our portfolio company behind Formae:

Open Jobs

Keycard is gaining serious traction as the control plane for autonomous agents, and they're growing the team to match. The problem in front of them: agents can't go to production without identity and access control, and nobody's shipped the answer yet.

If you want to help write that spec rather than wait for someone else to, they're hiring across Console, Data, Identity, SDK, and Systems, all remote. See open roles here.

Many of our other portfolio companies are hiring across similar roles, explore more here!

Events

Clarify Co-Founders are hosting a virtual demo today (7/16) at 9:30am PT. Patrick Thompson and Ondrej Hrebicek will show you what a day in an autonomous CRM looks like now that agents, Rep, and prospecting tools are all live in Clarify. If you’ve been curious what this looks like in practice, this is the session. Register here.

Workbrew is hosting a live webinar on July 23: Cask Vulnerability Reporting: Closing the Mac Fleet Security Gap. It's a walkthrough of their new feature that surfaces CVEs across Homebrew-installed GUI apps (casks) on your Mac fleet, aimed at IT/security teams who can audit CLI packages but have a blind spot on GUI installs. Register here.

Attending AI4 2026? Join MotherDuck CEO, Jordan Tigani, on Aug 4th at The Venetian for the Data Engineering panel: Building the Modern Data Stack for AI: What's Working and What's Not. RSVP here.

Thanks for reading! We'll be back next month with more updates.

— The Essence Team

Forwarded this email? Subscribe here

Asim Moinuddin, Tim Chen, Naomi Walker-Garrett

Keep Reading