How a Verification-First AI Harness Outperforms Bigger Models in Bug Repair

·1 views

A software engineering team has published a method for integrating a language model into automated maintenance pipelines while strictly limiting its decision-making authority. The approach, derived from a production deployment managing roughly 1,400 extraction targets and 1.6 million monitored items, centers on a machine-checkable oracle that verifies every AI-proposed fix before it is applied. The team found that constraining prompts and narrowing evidence roughly doubled precision, while doubling model size yielded no measurable improvement. The method is designed for problems where artifacts are small and structured, failures are high in volume, and incorrect repairs would produce silent wrong output rather than obvious errors. The authors present the framework as domain-independent, noting it has been validated beyond web extraction in configuration migration and flaky-test repair scenarios.

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)

Cloud AI Coding Agent Founders Read Every Message From All Six Early Users

A small team building a cloud AI coding agent manually reviewed every message sent by their first six users over 21 days, uncovering critical product flaws. A key bug caused the agent to silently report "nothing happened" after spending over two minutes and 254,000 tokens reading a codebase — because the progress tracker only counted file writes and shell commands, ignoring reads and searches. A separate issue sent users a blank white screen instead of a live preview due to a missing base path in scaffolded Vite apps — a fix that had been sitting in a branch but never pushed to production. On a more positive note, one user returned unprompted nearly 13 hours later to continue building after the agent generated a substantial application skeleton, with no email nudge or onboarding prompt. The team concluded that their biggest obstacle was not agent reliability alone, but their own inability to distinguish meaningful automated work from unproductive thrashing in their logs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Tool Compares A2A Agent Cards Across AWS, GCP, and Azure Clouds

A developer tutorial published on August 25, 2026, demonstrates how to fetch and compare Agent Card metadata from A2A-protocol agents running on three major cloud platforms: AWS Bedrock AgentCore, Google Cloud Run, and Azure Container Apps. An Agent Card is a JSON document hosted at a standard path that describes an agent's name, skills, and endpoint details, and is the first thing a client reads before connecting. The project only performs discovery — it never invokes the agents, sends prompts, or incurs any token costs. A key finding is that only AWS Bedrock AgentCore separates card-discovery permissions from invocation permissions via a distinct IAM action, while Cloud Run and Azure Container Apps bundle both behind a single access role. The accompanying open-source repository includes injectable test transports to simulate edge cases such as authentication denials and legacy card paths without requiring live cloud environments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Ripple Recommends Withdrawing XChainBridge Amendment XLS-38 After Low Adoption

Ripple has recommended that the XRP Ledger community withdraw the XChainBridge amendment (XLS-38) and the related fixXChainRewardRounding amendment, citing insufficient developer interest over a 12–15 month evaluation window. XLS-38 was originally designed to enable native cross-chain bridges on the XRPL, allowing assets to move between mainnet and custom sidechains via a decentralized network of witness servers. Ripple concluded that the witness server model posed difficult trade-offs between security, decentralization, and governance, particularly as bridge value scales. The company chose Axelar — a purpose-built bridging network with 75+ validators and experience across 55+ blockchains — to serve the XRPL EVM Sidechain instead, a decision announced in June 2024. With the EVM Sidechain bridge needs met by Axelar and no significant private sidechain projects emerging that required XLS-38, Ripple determined the amendment no longer warrants activation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

brainstorm-mcp v1.4.0 Adds API-Key-Free Multi-Model Debate via Host Agents

Developer tool brainstorm-mcp, which facilitates structured multi-model AI debates, released version 1.4.0 on March 24, 2026, introducing a 'hosted mode' that requires no external API keys. Previously, users running Claude Code had no usable keys because the server only supported non-Anthropic providers, creating unnecessary onboarding friction. In hosted mode, the server never calls any model API directly; instead, it constructs debate prompts and returns them to the host, which spawns its own sub-agents to generate responses. Mode selection is automatic based on whether model names include a provider prefix, requiring no additional configuration flags. The update also added context injection, multiple debate styles, and a structured synthesis format to close each debate session.

0 comments Read more at DEV Community