Why AI Coding Agents Need a Test-Driven Harness, Not Just Better Prompts

·3 views

The author argues that developers often rely on increasingly elaborate prompts to steer AI coding agents, even though agents working in projects without tests or clear constraints can still falsely report success and break existing code. The proposed alternative is harness engineering, which surrounds an agent with an environment of project documentation, failing tests, architectural guardrails, and a single verification command. The tutorial walks through building such a harness using Node.js with no paid APIs or npm dependencies, in about 15 minutes. Its example is a pricing function that wrongly accepts negative quantities, NaN prices, and discounts above 100%, which the new tests are designed to expose. The author frames prompts and harnesses as complementary: prompts state goals, while harnesses make progress observable and verifiable.

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)

Study of 1,355 App Store reviews finds slowness and pricing drive low ratings for a Reels blocker

An analysis of 1,355 App Store reviews for SocialLite, a Reels and Shorts blocking app for iOS, covered February 1 to September 6, 2026. Of these, 466 gave one to three stars, and the most frequent complaint in those ratings was that the app was slow, heavy or crashed, cited in 30.7% of them. Less common but strongly linked to low ratings were issues such as inability to switch between multiple accounts, Screen Time restrictions affecting unrelated websites, and login problems. Reviewers were also split on the free tier, with some objecting that suggested posts still appeared while others praised the free core features. The author, who builds a similar app, disclosed this interest and noted the findings apply only to this single app.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Codex CLI can queue bounded runs in osFoundry team and loop sessions as a Runner

A new guide explains how Codex CLI can act as a Runner inside osFoundry team or loop sessions. From the terminal, Codex queues a bounded objective with request_session_run, and the session's own agents carry out the work on its runtime. Codex then reads the outcome, leaves a note on the session's Team mail, and collects the agents' replies. The owner approves the connection once and assigns the Runner seat per session, so individual runs need no further clicks. Codex is configured to ask for confirmation before any write-capable tool runs, and the owner can remove the seat or revoke the connection at any time.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five Open-Source AI Agent Patterns for Production Deployment on Azure

Demos of AI agents often end at a simple prompt-and-response loop, while production systems also require current information, controlled permissions, cost-aware routing, dependable scheduling, and audit logging. An article on DEV Community highlights five open-source projects that add these capabilities using Azure AI Foundry, Azure Container Apps, and Azure Key Vault. Popularity figures cited are from a November 12, 2026 snapshot and indicate momentum rather than quality or security. The first project, Pinky, lets agents propose phone actions while requiring human approval for sensitive steps, with the author recommending narrow tools and nonce-bound approvals. The second, the Kagi Search API, can be wrapped in an MCP server to ground agents in fresh web evidence with credentials stored in Key Vault.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

JSLab Restores Tabs, Layout and AI Chats Across Relaunches

JSLab, a JavaScript and TypeScript scratchpad for macOS, keeps your work intact when you quit and reopen it. On relaunch it restores your tabs and their code, along with each tab's layout, folds, cursor and scroll position, plus the window size and display. Each tab's working directory, Web View state, and AI Chat conversation are also brought back. The session is stored in session.json, and if that file is unreadable, JSLab falls back to its backup and notifies the user. A tab whose contents cannot be read opens empty and read-only to prevent overwriting the file on disk.

0 comments Read more at DEV Community