AutoScout24 described a practical way to reduce Claude Code and Codex cost without changing the main workflow: remove excess from the agent environment before it reaches context. The problem is not only model price. The agent pays for every noisy log, blind file read, and verbose answer even when the fix needs one failing assertion or one function name.
What working pattern reduces agent cost?
Cost control has several layers:
- Trim: compress logs, test output, and service lists down to error, path, and exit code.
- Map: ask a code index first, then read specific files.
- Route: choose model, effort, and subagent by task risk, not habit.
- Cache: keep a stable prefix and avoid mixing dynamic data with durable rules.
Prompt caching supports this scheme but does not replace it. Anthropic caches the request prefix: system, tools, and messages up to a cache breakpoint. Reused context can be much cheaper, but the cache is fragile. Changing tool order, timestamps, or dynamic instructions can turn a repeat request into a cold read.
What risk does this create?
Cost savings can throw out diagnostics together with noise. If a hook drops a compiler warning, a code graph is stale after a refactor, or a cheap subagent misclassifies a task, savings turn into extra iterations.
Measure these settings not only by tokens saved, but by cost per completed task: green test, merged PR, or verified draft.
What should teams configure?
Cost control should live in team configuration: AGENTS/CLAUDE rules, hooks, MCP tools, model profiles, and output limits.
You cannot build a budget on the hope that every engineer will manually remind the agent to “read less” every time.