Research
Reducing Token Usage in Agentic Coding
Field notes on the tooling layer that keeps Claude Code from burning tokens — what I installed, what worked, and the candidates still on the bench.
Working document — updated as tools earn or lose their place in the stack.
Agentic coding has a silent cost curve: the agent reads everything, and you pay for everything it reads. Most of what flows through the context window is noise — full command output, entire files for a one-line answer, repo exploration that repeats every session. This document tracks the tools I use to cut that waste, and how they’ve held up.
Installed and proven
On 2025-05-26 I installed rtk and code-review-graph, and they work well.
- rtk — sits in the middle between the agent and the shell, filtering command I/O so only the important part reaches the model instead of the full dump. The hook-based setup rewrites commands transparently; the agent doesn’t need to know it’s there. Claimed 60–90% savings on dev operations, and my usage analytics are consistent with that.
- code-review-graph — builds a knowledge graph of the codebase so the agent can answer structural questions without re-reading the whole repo every session. The pitch that sold me: stop your agent reading the whole repo, which is the article I based my install on. Background reading: reduce LLM token consumption with code-review-graph.
On the bench
Candidates I’ve collected but not yet adopted:
- context-mode — plugin to improve context windows. Next in line to trial.
- graphify — the open-source alternative to code-review-graph: a self-updating knowledge graph for Claude Code. Worth watching if the hosted option ever becomes a constraint.
- Understand-Anything — broader comprehension tooling; unclear yet where it fits relative to the graph tools.
The autonomous angle
Token efficiency compounds with autonomy: commands that let the agent finish work while you sleep (goal, loop, batch) only make economic sense if each iteration is cheap. The filtering layer is what makes the loop affordable.
Open questions
- Where is the crossover point where a knowledge graph goes stale faster than it saves tokens?
- Does I/O filtering ever hide the detail that would have prevented a wrong turn? Haven’t caught it doing so yet — but I also wouldn’t have seen it.