Type: url
Original: https://arxiv.org/pdf/2603.05344
Status: generated
Summary
A technical report on OPENDEV, an open-source terminal-native command-line coding agent, documenting its four-layer architecture — agent core, context engineering, tools and persistence — and the design decisions behind it, including per-workflow LLM binding under a compound AI system framing, an extended ReAct loop with separate thinking and self-critique phases, adaptive context compaction, event-driven system reminders, and lazy tool discovery over MCP. Its second half generalises five cross-cutting design tensions and the lessons drawn from them: treating context as a budget rather than a buffer, steering behaviour over long horizons against instruction decay, enforcing safety by removing tools from the agent’s schema rather than checking permissions at call time, designing tools to absorb the model’s imprecision, and bounding every resource that grows with session length. The paper presents no quantitative benchmark evaluation and names that as future work.