Developer Finds Clean APIs Make AI Agents Lazier, Invents Friction-Based Fix

·1 views

A developer building Soma, an AI-powered codebase management system, achieved a 97.3% First Pass Success Rate by forcing the agent to delegate tasks to over 50 concurrent subagents via raw bash scripts. Migrating to cleaner MCP/JSON-RPC tooling caused performance to drop sharply to 87.5%, as the LLM abandoned delegation and began making sweeping code changes with minimal prior research. The developer identified this as the 'Abstraction Trap,' where easy-to-call APIs give LLMs false confidence and lead to context window saturation and erratic behavior. To counter this, he introduced Test-Time Compute Oracles — adversarial judges that intercept and scrutinize proposed changes before execution — alongside a just-in-time rule injection system to prevent context collapse. The experience led to a broader conclusion: agentic AI systems benefit from deliberate architectural friction rather than the clean abstractions that work well in traditional software engineering.

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)

GitLab Self-Managed Runs Two AI Code Reviewers Depending on User Seat Type

GitLab Self-Managed offers two distinct AI code review features under the same @GitLabDuo handle: the agentic Code Review Flow and the non-agentic GitLab Duo Code Review. Which feature runs is determined by whether the user who triggers the review holds a GitLab Duo Enterprise seat — if they do, the non-agentic single-pass review activates; if not, the more capable agentic flow runs instead. Code Review Flow can analyze repository structure and cross-file dependencies through multi-step reasoning, while GitLab Duo Code Review is limited to the merge request's file diffs in a single pass. This distinction is critical for teams enforcing standards that span multiple files, as the Enterprise-seat-triggered reviewer cannot read beyond the immediate diff. Group Owners can override the default behavior to pin all reviews to Code Review Flow regardless of seat type, with credit usage attributed to the initiating user.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Delegating Code to Claude Cuts Astra Tokens but Gains Are Inconsistent

A developer built a local tool called TaskRoute to route coding tasks from Astra in Codex to Claude Code, aiming to preserve Astra's token allowance for non-coding work. Early tests showed mixed results: one comparison found only a 3% reduction in Astra's uncached input tokens while taking 49% longer, and a smaller task actually increased token usage by 29%. A more controlled synthetic test on a new Python task showed stronger gains, with Astra's uncached input tokens dropping 59% and output tokens falling 83%, though elapsed time remained roughly unchanged. The author cautioned that Claude's separate token usage, failed runs, and session setup costs are not included in these figures, making direct subscription savings difficult to assess. TaskRoute remains a small experimental plugin for macOS, and the findings reflect limited sequential tests rather than a broad benchmark.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

FarmMemory AI Assistant Retains Field-Specific History to Help Farmers Track Crops

A developer has built FarmMemory, a voice-first AI assistant designed to help farmers recall field-level history across multiple conversations. The tool allows users to interact via text or voice in their preferred language, including Telugu, and links observations to specific field identifiers to prevent data from one field mixing with another. Built with Next.js, FastAPI, Groq, and a memory layer called Hindsight, the system explicitly separates memory storage from response generation. Only information from the farmer's own messages is saved as durable observations, ensuring AI-generated interpretations are never stored as factual farm history. The goal is to give agricultural AI assistants a reliable, inspectable memory without overloading language models with entire farm histories in every prompt.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

BrandBridge Dev Explains How to Build a React Frontend Before the Backend Exists

A five-person team is building BrandBridge, a marketplace designed to connect content creators, photographers, brands, and startups on a single platform. The frontend developer faced the challenge of building the React UI before any API contract was agreed upon with the backend engineer. To avoid a future rewrite, the developer drew a deliberate boundary between frontend and backend logic, structuring the app to function independently of a live backend. The platform supports four distinct user roles and includes an AI-driven content pipeline that learns from past performance to improve future campaigns. The frontend stack uses Vite, React Router, Tailwind CSS, and Recharts, with no state management library deemed necessary.

0 comments Read more at DEV Community