Maritime, a startup from Y Combinator's F26 batch, has launched a managed cloud built for AI agents that run in production. Co-founders Ary and Maria are pitching the platform as a full infrastructure layer, not just a place to run code. They announced the launch on Hacker News.
The founders argue that the current crop of sandbox companies, many of them based in San Francisco, addresses only a slice of the problem. Most of those providers buy bare metal servers from hosts such as Hetzner or OVH, use Firecracker to split the machines into virtual machines, and resell them at thin margins. Ary and Maria say that model is a narrow answer to a hard problem.
Maritime says it spoke with a couple hundred companies building AI agents, from personal assistants and coding tools to specialized healthcare agents. The founders say the most common pattern was that even after picking a sandbox provider, teams still had to build a good deal of additional infrastructure before their agents worked.
Three points of friction
The first problem the founders identified was agent lifecycle management. Teams have to build sleep and wake functions, persistent storage, triggers, scheduling, per-user routing, spend caps, health monitoring and failure recovery. They also need a way to push updates across thousands of virtual machines without taking services offline. Each task is small on its own, but the founders say they add up quickly.
The second issue is custom environments. Production agents often need very different setups. Some require Windows applications, others need Android, browsers with persistent sessions, custom operating system images, preinstalled desktop software or specific dependencies. Some agents must reach private networks to work with customer systems on premises, some need residential IP addresses, and others must be cut off from the network entirely.
The third is cost. The founders gave an example: a personal assistant sold for $20 a month would be squeezed by a sandbox that costs $0.04 an hour, which comes to roughly $29 a month if it runs continuously. That figure does not include large language model inference, storage or other expenses. Teams can reduce the cost by taking snapshots automatically, making workloads event-driven, mounting S3 buckets through FUSE for cheaper storage and moving older snapshots to colder tiers. Ary and Maria say those fixes also require more infrastructure to build and maintain.
How Maritime approaches the build
Maritime says it gives each agent its own isolated, persistent microVM with a configurable environment. On top of that compute, the company has built a set of services that include provisioning, persistent state, sleep and wake orchestration, triggers, schedule creation, monitoring, network configuration, updates, rollbacks, recovery, per-agent spend tracking and login handoffs.
The founders also said they put heavy effort into cost. Their work includes scale-to-zero, copy-on-write over shared base images, compressed memory snapshots, and hot, warm and cold tiering for snapshots and disks. They say they balanced those changes against performance measures such as latency.
The company's pricing is one of its main selling points. Maritime charges $1 per month for a 2GB RAM agent, with unlimited usage included at that price. Additional resources are available as needed.
Ary and Maria have research backgrounds that shaped the company. Maria was a visiting lecturer at MIT and did research at Project NANDA focused on agent infrastructure. Ary researched agent coordination protocols at the same lab. The two started Maritime together this March, after running into difficulty deploying agents to the cloud during their research.
The launch post is also a request for feedback. The founders want to hear from developers about what they had to build on top of sandboxes to get agents working, along with any workarounds and odd solutions. They also asked whether there are projects that developers tried to build on virtual machines but could not finish, whether Maritime fits their needs, and what feature might make the platform more appealing. The founders said they would answer questions in the comments.
Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.