Launch HN: Agent37 (YC F26) – Low-cost agent sandboxes for running agents
Agent37 launched its low-cost sandboxes for running AI agents, boasting prices significantly below competitors thanks to shrewd infrastructure choices and focused optimization. The Hacker News community is abuzz with both congratulations and critical questions about the long-term viability of a price-centric strategy in a rapidly commoditizing market. This story highlights the intense competition and innovative approaches emerging in the AI infrastructure space.
The Lowdown
Agent37, a YC F26 startup, has unveiled its platform offering low-cost sandboxes specifically designed for running AI agents. The company claims an average cost of $0.37/month per sandbox, with a free plan made possible by aggressive cost optimizations.
Key aspects of their approach and offerings include:
- Cost-Efficiency: They've significantly reduced costs by analyzing the bursty CPU/RAM usage patterns of agent harnesses, rather than over-provisioning.
- Infrastructure Strategy: Agent37 avoids expensive hyperscalers, opting for providers like Hetzner and OVH, which they identify as a major cost-reduction lever.
- Technical Stack: They leverage gVisor for better utilization and faster boot times, sharing host memory across sandboxes, though this comes with trade-offs like not supporting Docker-in-Docker.
- Hardware Optimization: Servers are procured with a focus on maximizing RAM, as it's typically the cost bottleneck for agent workloads.
- Background: The founder, Vishnu, brings experience from Amazon (where he built internal AI chatbot Diya and coding harness Ziya) and Quantcast.
- Target Audience: Agent37 is actively seeking developers building messaging agents to test their platform's limits.
By niching down on agent-specific use cases and meticulously optimizing for cost-performance, Agent37 aims to provide an economical solution for the burgeoning AI agent development ecosystem.
The Gossip
Competitive Crucible & Pricing Predicament
The discussion heavily scrutinizes Agent37's strategy of competing primarily on price in what many perceive as a commoditized and fiercely competitive market. Commenters express skepticism about the long-term viability of a 'race to the bottom' against well-funded incumbents or other YC startups in the same space. The founder counters that this pricing is a 'wedge' strategy to gain market share before differentiating further, while acknowledging the intense competition and high number of similar startups within their YC batch.
Sandbox Scrutiny & Self-Hosting Solutions
Users are keen to compare Agent37 with existing and alternative sandbox providers. Direct comparisons are drawn to services like Boat.dev, Fly.io's Sprites, and Maritime.sh, often focusing on pricing, storage, and specific feature sets. Additionally, a significant portion of the conversation explores the concept of self-hosting AI agent playgrounds on local Linux machines, with users discussing tools like Docker, Podman, Tailscale, and Google's Agent Substrate, and lamenting the lack of comprehensive open-source solutions for such setups.
Practicalities & Puzzling Provisions
Commenters dive into the practical aspects and user experience of Agent37's service. Questions are raised about the specific LLM used in the free tier, the technical architecture of the 'harnesses' (agentic loops), and unexpected limitations like URL fetching quotas. Users also share their personal use cases, such as offloading Android emulation, and offer feedback on areas like simplifying Google API integrations, highlighting common pain points in agent development environments.