Token Efficiency – Your Guide To Smart AI Development

A Devtoberfest 2026 session from two engineers on SAP's business AI platform team on where a coding agent's tokens go: the read-decide-act loop behind one request, what the context command shows is occupying the window — 22,700 tokens of tool definitions against 1,800 of project instructions in their example — and the habits of compacting with a hint, clearing between tasks, choosing model and reasoning effort per job, and handing a large investigation to a sub-agent so only its summary comes back. It reaches agent guidanceAgent guidancecapabilityPackaged knowledge you hand a coding agent so it gets a framework right unprompted — skills, rule files, instruction packs. Distinct from agent tooling, which is about building agentic features into your application, and from agent access to your code, where the agent reads the real project instead of being handed a file about it.SAP and community, core adopted it through the part on skills: a skill holds the reusable instructions for one kind of work and is disclosed progressively, a short description in the window and the detailed file read when it is relevant, rather than every occasional procedure sitting in the always-loaded project instructions. The demo runs on Claude Code against a generated warehouse-management project, with a custom status line showing context, session cost and a self-set budget.