Symbolic AI is making a comeback. Instead of relying purely on neural networks, symbolic AI represents knowledge through rules, logic, and structured relationships. That means systems that can reason step-by-step, produce verifiable outputs, and be understood better by humans.

Probabilistic AI systems are often brittle in production and they fail unpredictably when tasks require consistency, multi-step reasoning, or strict correctness. Symbolic approaches introduce deterministic structure, separating planning, reasoning, and execution into components you can inspect, test, and control.

A new generation of companies is building here. OpenSymbolicAI combines LLMs with symbolic primitives and deterministic execution, treating plans and workflows as structured programs rather than prompt chains. The core idea is that LLMs handle intent interpretation while symbolic systems ensure consistent, auditable outcomes.

We're seeing this in vertical applications too. In our portfolio, Fearn.ai applies symbolic AI to patents, structuring technical knowledge and legal logic into representations that can reason over prior art and claims far more rigorously than statistical approaches alone. In domains where a single hallucination has material consequences, this matters.

These developments point toward a broader shift from systems that are merely predictive to systems that can reason with structure and accountability. We believe symbolic and neuro-symbolic approaches will play a foundational role in what comes next.

Streamfold was recently acquired by Cursor 🎉 Huge congrats to Ray, Mike, and the rest of the team. Read more here.

Butter was recently acquired by Modal, keeping it in the Essence family 🤝 Congrats to Erik, Raymond, and the rest of the team. Read more here.

OpenInfer responded to Anthropic’s limitations of Claude use for OpenClaw by releasing an infrastructure routing layer that automatically directs AI/agentic workloads to the cheapest or most available compute resources, combating vendor lock-in from model providers. Check it out here!

Letta released the Letta Code app, built for memory-first agents that improve with use. Read more here, or try it out!

Cline released Kanban, a standalone app for CLI-agnostic multi-agent orchestration. Claude and Codex compatible. Check out the launch or give it a spin!

DuckDB launched a Claude Code plugin that adds DuckDB-powered skills for data exploration and session memory. Read about it here or take it for a spin!

Here’s our most recent podcast episodes from The Infra Pod with Catherine Jue from Kernel and the Open Source Startups Podcast with Pedram Amini from Maestro.

Speaking of Modal’s acquisition of Butter earlier, they are hiring!

If you’re not a fit for Modal’s current openings but are passionate about joining innovative AI and software infra startups, check out other exciting opportunities within Essence’s Portfolio.

For the second installment of our GTMfor.Dev reboot, we're sitting down with Ryan Blue — the engineer who shipped Apache Iceberg at Netflix, co-founded Tabular to commercialize it, and navigated a landmark acquisition by Databricks. If you're a technical founder working through the gap between great product and real traction, this one's for you. Save your virtual seat for this live podcast here.

Essence is also co-hosting an AI Infra After Dark Social in NYC on April 22nd at Ms. Yoo’s. We're bringing together people working across the stack: agents, model tooling, data systems, evals, observability, and everything it actually takes to run AI in production. Joining us are MongoDB Ventures and Vermilion Cliffs Ventures. If you’re building the future of AI infrastructure, we’d love to see you there.

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

— The Essence Team