Coding agents are being tuned with repo setup, deterministic flow, and shared code vocabulary

agent-stack treats token cost as a repository setup problem. Its command generates Claude Code and Cursor files, writes CLAUDE.md, AGENTS.md, .claudeignore, skills, hooks, Cursor rules, and a code map. The code map gives the agent one compact symbol index before it opens source files.

The README claims a setup under two minutes, 20 generated files, two wired hooks, and a verified CLAUDE.md capped at 800 startup tokens. Its example reports 7,180 current input tokens per day against a 12,340 baseline, a 41.8% drop. These are project claims, not an independent benchmark, but the measurement hook and stored baseline make the cost target explicit.

BotCircuits separates process routing from model reasoning. Workflows live as JSON files under .botcircuits/workflows/; after build, each workflow becomes a callable tool. The LLM handles each agentAction, while the runtime follows start, next, and compiled branch conditions.

The concrete promise is narrower than the title suggests. The README gives architecture and usage detail, including provider setup for Anthropic, OpenAI, and Gemini, a FastAPI gateway, and messaging channels. It does not report token-cost, latency, task-success, or deviation-rate results.

Martin Fowler’s essay gives the design-side caution for generated code. It defines code as both machine instructions and a conceptual model of the domain. The useful parts for LLM work are names, boundaries, invariants, and tests that humans and tools can share.

The essay has no quantitative results. Its value in this period is the maintenance argument: generated code can create cognitive debt when teams accept vocabulary and structure they do not understand. Stable abstractions and tests are presented as the context that makes LLM output easier to review and maintain.