Coding Agent Operating Controls

Teams using cheaper coding agents can make module boundaries executable. ANMA’s pattern is plain: developers write YAML contracts for each module, then anma sync generates Claude Code guidance, edit-blocking hooks, backend configs, CI checks, and optional CODEOWNERS entries.

The useful adoption test is a small repo with known dependency rules. Run the same boundary-sensitive task with a cheaper model on a plain checkout and on a checkout with generated guidance plus anma check in CI. ANMA reports that Claude Haiku 4.5 violated a Python boundary in 13 of 19 plain-repo runs and 0 of 20 ANMA runs. A TypeScript follow-up reports 18 violations in 20 control runs and 0 of 20 with ANMA. The authors also say Claude Opus 4.8 respected the Python boundary without ANMA, so the practical target is cost-sensitive agent use and CI governance for human or agent-written diffs.

Coding-agent spend can be checked at the session and turn level before it becomes a finance surprise. Lupen reads local Claude Code and Codex JSONL logs, groups activity by session, turn, step, skill group, and sub-agent, then recomputes costs from token counts and public price tables.

A practical workflow is to run a local report at the end of each agent-heavy day and add lupen budget --over 20 --last 7d or lupen verify to scripts used by power users. Both checks can exit with code 4, which gives teams a simple way to catch cost drift or runaway sessions without uploading prompts, file paths, images, or URLs to a hosted service. The evidence is feature-level rather than a benchmark, so teams should validate it against their own provider invoices and raw logs during a short pilot.

Teams can test whether an agent has learned an unfamiliar corpus before giving it open-ended work. Machine Studying defines this as a pre-task study phase over a document corpus, followed by hidden downstream questions. StudyBench applies the setup to DSPy code, OpenClaw code, and recent machine-learning literature, and its metric rewards accuracy at lower inference-token budgets.

The practical version is a small internal exam for a private repo, product manual, or research library. Give the agent time to study the corpus using its normal tools, keep the exam hidden, then score answers under fixed token budgets. The early result warns against assuming that search access solves the problem: Qwen3.5-9B on DSPy improved from 9.6 to 29.4 lenient score only when forced to use 20 search iterations, and the reported fine-tuning baselines did not reliably raise agent expertise.