Developer sandboxes AI coding agent in a dedicated VM to prevent accidental damage

·1 views

A developer running AI coding agents with shell access built a dedicated Linux VM on Apple Silicon to isolate the agent from his real files and credentials. The setup uses a snapshot taken before each session, allowing any destructive action to be rolled back in about 30 seconds rather than debugged. Building the sandbox on Apple's Virtualization.framework came with several non-obvious pitfalls, including the loss of a critical EFI variable store file that causes Linux guests to fail silently on reboot. Attaching a serial console for debugging also backfired, as Ubuntu's installer detected it and redirected the entire installation UI away from the display. The author concludes that truly isolating an agent means avoiding shared filesystems entirely, since mounting a real project folder into the VM defeats the purpose of the sandbox.

Read the full story at DEV Community

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

Discussion (0)

Rails 8 Upgrade Cuts Response Times 18% and Eliminates Redis Dependency

A development team migrated a mid-sized application from Rails 7 to Rails 8 and documented concrete performance improvements across multiple areas. Rails 8's Solid Queue, Solid Cache, and Solid Cable allowed the team to replace Redis and Sidekiq with database-backed alternatives, reducing infrastructure costs and operational complexity. The upgrade delivered roughly 18% faster average response times, 25% higher background job throughput, and a 15% reduction in CI pipeline duration. Asset pipeline changes via Propshaft and import maps also cut deployment compile times, while tighter Kamal 2 integration simplified zero-downtime releases. The team noted the migration required significant effort, especially around Solid Queue adoption and deprecated API audits, but concluded the gains justified the investment.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Concerns Grow Over US Fiscal Solvency Amid Rising Debt Worries

Growing concerns about the United States' long-term fiscal health have been gaining attention among investors and analysts. Questions are being raised about America's ability to manage its mounting national debt obligations. The discussion has surfaced in financial media, including coverage by the Financial Times. Analysts are examining whether current debt trajectories pose a credible risk to US solvency. The topic has sparked debate among economists and market observers about the country's financial stability.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Developer Builds AI-Readable Knowledge Base Using Obsidian and Git, No RAG Pipeline Needed

A developer at NEXT4I built a structured knowledge-base system using Obsidian, Git, and VS Code to solve the problem of documentation scattered across Google Docs, Trello, Apple Notes, and multiple code repositories. The system relies entirely on plain Markdown files organized into folders covering infrastructure, platform design, principles, and other domains. Because an Obsidian vault is simply a folder of .md files, the VS Code AI agent can read it natively without any custom integration, vector database, or embedding pipeline. Version control is handled by initializing a Git repository inside the vault, enabling full decision history and audit trails via a private GitHub repo. The approach was guided by three criteria: zero vendor lock-in, offline and online accessibility, and AI-readability without additional infrastructure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Verbose Agent Logs Are Not Audit Trails Without Evidence Contracts and Tests

AI agent runtimes can generate thousands of log lines yet still fail to answer what actions were actually taken after a failure, making true audit trails distinct from verbose logging. A reliable audit trail requires append-only events with stable run IDs, monotonically increasing sequence numbers, and a hash chain that makes tampering or omission detectable. Sensitive data such as tokens, credentials, and raw prompts must be redacted at write time rather than filtered later in a dashboard query. Each externally visible tool call should progress through defined states — from intent recorded through dispatch to confirmed or unknown outcome — with stable request keys to prevent duplicate effects during retries. The test suite must actively mutate the ledger and confirm the verifier rejects deleted, reordered, or forged records, ensuring the system fails closed rather than silently accepting corrupted evidence.

0 comments Read more at DEV Community