AI Agents
Browse all articles, tutorials, and guides about AI Agents
Posts
We Tested AWS Emulator: It Blocks More Than Its Docs Say, and Less Than AWS Does
AWS Emulator passed all 40 of our workflow checks and caught two IAM mistakes in our own test code. It also let a user with no policies assume an admin role, ignored permissions boundaries and session policies, and with the default JavaScript SDK settings, one SQS request went to real AWS. Here is what we measured on a throwaway droplet.
Who's On Call for the AI Agent? A Production Runbook
An agent that can call production tools is a service with an incident path, not a clever prompt. Assign an owner, page only for harm in progress, test a kill switch, scope its identity, and keep enough evidence to explain every action.
We Gave an AI Agent a read_only Neon Role. Here Is What It Blocked
A Neon read_only role blocked every INSERT, UPDATE, DELETE and schema change we sent, and the Neon MCP Server with a pinned role stopped every durable write. What still ran comes from standard Postgres defaults, and the real risk was a read that stalled a migration.
Build an Everyday AI Assistant with DigitalOcean Inference
Build Daylight, a personal assistant that turns a brain dump into a realistic plan, adapts when your day changes, and saves the changes you accept.
DigitalOcean Agent Droplets: When the $50 Plan Pays Off
Agent Droplets is a prepaid discount on DigitalOcean Managed Agents, not a server. We priced our real agent runs from a real invoice to find where the $50 and $200 plans start to save money, and found a rounding rule that made our runs cost more than the token math.
A Poisoned .git/config Runs Code on git status. We Tested Which Commands and Copies Carry It
On October 2 GitLab disclosed ConfigPoisoning: a repo that brings its own .git/config makes an AI coding tool run attacker commands when it shows a diff. We ran the same trick against plain git 2.39 and 2.55. A bare git status ran repo-supplied programs, the usual safe diff flags missed the clean filter, git 2.54 config hooks walked past core.hooksPath=/dev/null, a clone was clean but a cached workspace was not, and GitHub-hosted runners switch off the ownership check that would stop it.
A Repo per Agent: What Cloudflare Artifacts Changes About Git
Cloudflare Artifacts gives every agent, session or task its own Git repository. We measured how push contention grows when agents share one branch, walked through the Workers API, and priced the pattern so you can decide where it fits.
Issue to Pull Request with DigitalOcean Managed Agents
We gave an OpenCode agent on DigitalOcean Managed Agents three real GitHub issues and no GitHub token. It produced three merged pull requests and one wrong one that passed its tests. Here is the setup, the recorded runs, the cost per issue, and every gotcha we hit.
Where to Run AI Agents: 8 Managed Agent Runtimes Compared
DigitalOcean Managed Agents, Cloudflare, AWS AgentCore, Google, Microsoft Foundry, E2B, Vercel and Modal, compared on isolation, state, tool access, limits and price, with one cost scenario worked out on every platform.
One Key for Claude, GPT, and Gemini: the Gateway Pattern
Using three model providers usually means three API keys, three SDKs, and three billing relationships sprayed across your code. An AI gateway collapses that to one credential and one OpenAI-compatible endpoint. I proved it on a Neon Function: the same call answered by GPT, Claude, and Gemini.
A Postgres-Backed MCP Server in ~20 Lines
Most of what an MCP server does is run database queries on behalf of an AI agent. So I put one right next to the database. Here is a Postgres-backed MCP server built on Neon Functions, deployed onto a database branch, with the code, a live client test, and the repo.
Streaming an AI Agent Without a Function Timeout
Long agent loops and long token streams run into the same wall: a serverless function that hits its execution cap and cuts the connection. Neon Functions hold long-lived streaming connections by default. I deployed two endpoints to prove it: one streamed for 90 seconds, the other streamed an agent token by token starting at 466 ms.
I Gave an AI Agent a Database, Compute, Storage, and Models From One CLI
An AI agent usually needs four accounts: a database, somewhere to run, object storage, and a model provider. I wired all four from a single Neon credential and had a deployed image-generating agent in a few minutes. Here is the actual build log, the config that ties it together, and the honest caveats.