Developer builds multi-agent AI coding orchestrator to keep humans in the merge loop

·1 views

A developer created agents-concerto, a multi-agent orchestration system built on top of Claude Code, after growing frustrated with opaque AI coding sessions that produced untraceable changes and meaningless green tests. The system uses a conductor-and-players model where separate agents handle implementation, classification, and review, with strict rules preventing any code-writing agent from also approving or merging changes. All acceptance criteria are written in Given-When-Then format at the planning stage and must translate into tests that assert only observable behavior, never internal implementation details. Each agent operates in its own isolated git worktree, and the system caps review-fix cycles with a human exit option once the limit is hit. The orchestrator's sole output is a pull request ready for human review, with auto-merge explicitly excluded and merge permissions never requested.

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)

Developer Builds Autonomous AI Agent on Solana With On-Chain Policy Controls

A developer participating in the #100DaysOfSolana challenge has built an autonomous AI agent capable of reasoning, making decisions, and executing transactions on Solana's Devnet. The system was assembled over five days, with each session contributing a distinct component: an agent loop, blockchain tools, an MCP server, a policy engine, and a final autonomous workflow. The agent uses a large language model to decide which actions to take, but all fund transfers must pass through a deny-by-default policy engine before any transaction is signed. The developer noted that the policy engine, rather than the AI model, serves as the true core of the system, since the LLM only determines intent while the surrounding infrastructure enforces safety. The project culminated in a documentation effort aimed at making the entire architecture reproducible by any developer without additional guidance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Senior Engineers Read Tickets Before Writing a Single Line of Code

A fictional dialogue series on DEV Community uses an uncle-nephew conversation to explore how software is built from idea to production. In the first episode, an experienced engineer challenges his nephew — who has just completed React and RAG courses — to reconsider what separates junior from senior developers. The uncle argues that writing code is the final step of engineering, not the first, comparing it to a civil engineer who studies a river thoroughly before laying a single brick. Using a simple 'Add Wishlist Feature' ticket as a case study, he demonstrates how plain-language requirements hide dozens of unresolved decisions around authentication, data persistence, and edge cases. The key lesson is that engineers must surface ambiguities that affect database schemas or API contracts before coding, while making reasonable assumptions about minor details to keep the sprint moving.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Auditable Palworld Breeding Calculator With Versioned, Verified Data

A developer has built an independent, ad-free Palworld breeding calculator called Palworld Breeding Combos, designed around data transparency and reproducibility. Each generated dataset carries a manifest recording the game version, source revision, generation timestamp, and verification status, so users always know exactly what data produced a given result. Rather than reimplementing the breeding formula, the project imports a pinned public dataset from a specific upstream release, normalizes it, and excludes any values that cannot be confidently mapped instead of guessing. A known data gap involving the Pal 'Astralym' was handled as a separately documented overlay, accepted only after two independent public sources confirmed the missing relationship. Validation scripts confirmed 288 of 288 breeding ranks and 251 of 251 unique combinations matched independent reference data, with seven unresolved codes excluded rather than approximated.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Open-Source Tool to Make LLM Conversation Context Manually Editable

A developer has created ThoughtDAG, an open-source prototype that reimagines how users interact with large language models by representing conversations as a directed acyclic graph on an infinite canvas. Each question-and-answer exchange becomes a node, and the edges between nodes directly determine which prior exchanges are included in the model's next request. Users can branch, merge, or prune conversation paths, giving them explicit control over what the model 'remembers' at any point. The tool runs entirely in the browser using IndexedDB, requires no account or hosted database, and supports both local inference via Ollama and OpenAI-compatible endpoints. Built with React, TypeScript, and the Vercel AI SDK under an MIT license, the project raises the open question of whether users prefer manual context control over automated memory management by agents or retrieval systems.

0 comments Read more at DEV Community