From Chat Snippets to Embedded AI Coding Agents
Engineering teams are rapidly moving beyond copy‑pasting snippets from browser chats into their editors. ClickHouse describes three levels of AI‑assisted development: basic chat use, integrated agents inside the CLI or IDE, and experimental autonomous agents running in isolated environments. While level‑one chat remains useful for quick exploration, it is increasingly obsolete for serious work on large codebases. The biggest productivity gains now come from level‑two agents that can read repositories, run commands, edit files, build, test, and even commit changes under human supervision. ClickHouse engineers report that the line between “use the agent” and “don’t bother” shifted over the past year as tools matured, especially with advances such as Claude Opus 4.5. Their experience suggests that the practical future of AI coding agents lies in tight integration with everyday workflows, not in loosely connected chat windows or fully autonomous black boxes.
ClickHouse’s C++ Codebase: Where AI Agents Actually Help
ClickHouse’s year‑long experiment using AI coding agents on its main C++ repository highlights concrete productivity wins. Early in the rollout, agents handled JavaScript boilerplate and small Python scripts but struggled with the complex C++ codebase. With newer models, engineers began delegating tightly specified C++ tasks, then bug investigations from CI logs, and eventually small features. The results were consistently better than expected. Patterns emerged: agents excel at repetitive build‑system adjustments, multi‑file configuration edits, and infrastructure boilerplate such as Kubernetes manifests, where they make fewer mistakes and never lose focus. They also resolve merge conflicts more accurately than humans in almost all cases, especially when paired with a “agent does, you review” workflow. However, long autonomous loops still deliver unreliable outcomes, underscoring that human oversight and well‑scoped tasks remain essential for safe, large codebase automation in production environments.
Cost‑Aware Agent Design: Reasonix and Cache‑First Workflows
As AI coding agents become viable for daily development, their runtime cost has turned into a serious design constraint. Reasonix, an open‑source DeepSeek‑native terminal coding agent, targets this issue with a cache‑first architecture. Long shell sessions traditionally resend the same repository context and instructions to the model, inflating bills. Reasonix instead leans on DeepSeek prefix caching to reuse shared context across turns, cutting repeated processing when developers stay within a single workflow. Its launch pitch cites a one‑day study claiming costs of about USD 12 (approx. RM55.20) instead of about USD 61 (approx. RM280.60) under a specific setup, though that evidence remains project‑published. The tool also emphasizes plan mode and first‑class Model Context Protocol support so agents can reliably connect to external tools and data. Overall, it reflects a broader shift: serious AI coding agents now compete not only on capabilities, but on their ability to keep long‑running sessions economically sustainable.
Why Terminal‑Native Agents Are Challenging IDE‑First Workflows
Reasonix’s launch as a terminal‑first agent signals that the shell is once again a primary battleground for developer productivity. Earlier tools like CLI‑based assistants and built‑in MCP servers already nudged AI‑driven work back toward the command line, and Reasonix doubles down on that trend. It runs on macOS, Linux, and Windows, requires Node.js 22 or later, and targets developers who are comfortable living inside local shells and npm tooling. By connecting via the Model Context Protocol, it avoids being trapped in simple prompt‑response loops and instead orchestrates tools and data sources much like richer IDE integrations do. For teams already standardizing on terminal workflows—scripts, CI runs, container tooling—a terminal coding agent becomes a natural extension that can read repositories, make multi‑file changes, and execute commands without forcing a specific editor. Together with ClickHouse’s experience, this points to a future where terminal‑native agents stand as credible peers to IDE plugins for serious large codebase automation.