SShortSingh.
Back to feed

Maritime Launches Managed Cloud Platform for Running Production AI Agents

0
·1 views

Maritime, a Y Combinator-backed startup co-founded by Ary and Maria, has launched an agent-native cloud for production AI agents. The founders argue that sandbox providers address only a small part of the infrastructure challenge, since developers still must build lifecycle management, custom environments, and cost controls themselves. Based on conversations with a couple hundred companies building agents, they identified friction points including sleep/wake handling, persistent storage, failure recovery, varied OS and network requirements, and high costs from continuously running sandboxes. Maritime gives each agent its own isolated, persistent microVM along with composable services such as monitoring, updates, rollbacks, and per-agent spend tracking. The company says it prices a 2GB RAM agent at $1 per month with unlimited usage, plus extra resources as needed.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Lakegres Enables Vector Search on Existing Open Table Format Data Lakes

Onehouse has introduced Lakegres, a tool that brings vector search capability to tables stored in open table formats such as Apache Hudi, Apache Iceberg, and Delta-style lakehouse formats. The approach lets teams run similarity search directly on data already sitting in their lakehouse, without copying it into a separate vector database. The announcement was shared on Hacker News, where it drew modest early interest with 7 points and 2 comments. The article describes the product as a way to simplify AI and semantic search workloads over existing analytics data. Full technical details are available in Onehouse's blog post.

0
ProgrammingDEV Community ·

Map Every AI Agent Action and Gate the Irreversible Ones Behind Human Sign-Off

The article argues that AI agents able to issue refunds, cancel orders, send emails, or move money pose a greater risk than those that only draft replies, because their results can be impossible or hard to reverse. It recommends listing every concrete action an agent can take, then asking whether each can be quickly and fully reversed. Actions that cannot be undone, such as sent messages, submitted payments, or taken-effect cancellations, should require human approval. Organizations should name specific approvers with backups, set safe defaults when reviews stall, and log all irreversible actions. The author also advises testing whether the agent can bypass these controls before customers interact with it.

0
ProgrammingDEV Community ·

DeOldify, DDColor and BigColor Compared on Colorization Error Using CIEDE2000

On ten grayscale-converted Prokudin-Gorskii photos, DeOldify's median whole-image color error was 9.55 ΔE2000, close to DDColor's 9.32 and BigColor's 9.04. Its results were less saturated, with mean chroma of 9.5 across 34 measured zones versus 18.9 in the original photos. The model also missed skies in some images, and the hosted version added a small icon in one corner. The test used Replicate with default settings, and errors were computed per pixel in CIELAB with a regional breakdown. The DeOldify repository was archived in October 2024.

0
ProgrammingDEV Community ·

How Opening Two Breakers on a Substation Ring Bus Splits the Network

A ring bus links circuit positions into a loop so power can flow either way around it. With all six breakers closed, the ring operates as a single electrical bus, and opening one breaker still leaves a connected path. Opening two breakers can cut the ring into separate sections, each of which must be analyzed independently. A section cut off from both incoming circuits has no power source, so no solved result should be reported for it. The article uses a 230 kV example with four demands to illustrate these effects.