Software engineering tooling has long operated on predictable flat-rate subscriptions: an IDE license costs the exact same whether an engineer writes one test or pushes fifty commits before lunch. As Lydia Hallie highlighted in Anthropic's operational guide on Claude Code, agentic development breaks that predictable subscription model wide open. In an agentic terminal workflow, every single task carries its own direct compute bill determined entirely by session hygiene.
Consider the practical delta: one engineer running Claude Code resolves a failing test in three surgical turns by targeting the exact file with an `@-mention`, while another lets the agent wander through unguided repository searches, pulling dozens of extraneous files into the conversation history.
"Being efficient with tokens doesn't mean using fewer of them overall. It means making sure the ones you do use go towards the thing you actually asked for."
Both workflows produce the identical git diff, but the undisciplined session consumes orders of magnitude more GPU capacity and forces the model to reason through massive context bloat on every subsequent turn. Engineering leaders need to treat inference billing as physical hardware allocation. Input tokens processed during prefill—including system prompts, `CLAUDE.md`, conversation history, and raw command stdout—re-run through compute on every prompt. Meanwhile, sequential decode output tokens (such as extended thinking and tool call payloads) hold GPU hardware active for significantly longer, driving up operational costs accordingly.
Technical Levers for Prompt Caching and Context Control
Controlling session spend requires active prompt cache management and aggressive pruning of the conversation surface. Prompt caching relies on strict prefix matching across consecutive requests; changing the target model or tweaking the reasoning effort level mid-stream invalidates the cache instantly and forces expensive full-context recomputations. Tech leads must enforce setting baseline parameters before launching tasks.
Furthermore, because Anthropic prompt caches carry a one-hour TTL, running `/compact` before stepping away from the keyboard ensures conversation summarization occurs while the session state is still cached at lower cost.
Context hygiene also determines how engineers feed raw data to the agent. Using direct `@-mentions` attaches target files immediately, bypassing expensive tool-driven search iterations and redundant Read calls. Terminal outputs represent another silent cost driver: unmanaged build outputs become permanent baggage in the conversation state. Engineering managers should train developers to apply quiet flags to verbose commands or offload noisy operations to isolated subagents.
Running `/context` at the start of a session audits loaded MCP tool definitions and `CLAUDE.md` instructions, exposing dead weight before a single token is billed, while invoking `/clear` between distinct tasks halts context pollution before it degrades reasoning quality.
For engineering leadership scaling AI-native developer workflows, token consumption is a direct function of operational discipline. Establishing baseline rules around session resets, cache retention, and subagent isolation converts Claude Code from an unpredictable budget black hole into a precise, high-leverage development environment.