The tax on good behavior.

There is a lazy critique of agentic AI systems that goes: they're bloated, they burn context on nothing, the overhead is waste. Someone captured Claude Code's first API request off the wire — the actual JSON that leaves the client before a single token of your problem gets processed — and the temptation is to read the resulting numbers as an indictment. ~18,000 tokens, ~73KB, before the model has seen your repository, your question, or your intent. Three-quarters of that is tool definitions the model may never call.

That is the wrong reading. Sit with the data long enough and it inverts: the tax isn't waste, it's the price tag on the restraint you actually experience when you use the tool. What follows is Agent Adda's reverse-engineering pass on that captured request — and the point of view that falls out of it.

The exercise intercepted a live Anthropic-compatible POST /v1/messages call — the literal payload Claude Code sends before doing anything.

Fig. 1 — Startup request anatomy (v2.1.260)

The first finding that matters: your project isn't in here. No repo contents, no bulk-loaded source, no README dump. Files load on demand, via Read / Glob / Grep, only when the model decides it needs them. What's sitting in that 18K-token floor is capability, not content — the standing inventory of things Claude could do, priced whether or not it does any of them.

Fig. 2 — Repo source is not in the floor; it arrives only via tool results

Within tool definitions, prose descriptions outweigh machine-readable schema by roughly two to one — 62.6% description text against 33% schema JSON. That ratio concentrates in orchestration tools.

Fig. 3 — Tool-family weight in bytes (approx.)

Five tools, roughly 20KB of prose between them — most of it not documenting a parameter. Core coding tools are not the expensive ones.

Here is the opening clause of the Workflow tool's description, in spirit: only call this when the user has explicitly opted in — a specific keyword, a direct ask in the user's own words, or a skill that tells you to. For any other task, even one that would clearly benefit from parallelism, do not call this tool.

ScheduleWakeup tells the model not to poll for work the harness will notify anyway, and to respect cache-window economics when sleeping. CronCreate tells it to avoid :00 and :30 because every other session making the same "reasonable" choice would thundering-herd the API — and to nudge off that mark even when the user asked for a round number.

Fig. 4 — Guardrail-as-affordance: constraint travels with the capability

The visible behavior — an agent that doesn't fork twelve subagents for a one-line question, that picks 8:57 instead of 9:00 for a digest — isn't emergent politeness. It's paid for, in tokens, on every turn. The elegance is a recurring cost, not a one-time training artifact.

Tools dominate, but the remaining quarter is not empty. System blocks (~9%, ~1.6k tokens) break into a billing header, identity one-liner, and ~6.5KB interactive policy — where Memory + Harness alone are ~60% of that text. Messages (~15%) include your prompt plus an injected agent/skill catalog (~7.4KB / ~1.85k tokens) — sixteen named types the Agent / Skill tools can reach.

Cost hygiene levers are real: --bare, short CLAUDE.md, prune MCP, lean memory. Full vs bare is stark:

Fig. 5 — Full vs --bare startup floor

But deferring full tool descriptions until first use isn't merely theoretical — it's shipping. Cron control, cross-session messaging, worktree management, most MCP connectors sit as name-only stubs until a discovery call pulls the full schema.

Fig. 6 — Pay the lean core every turn; pay orchestration when you reach for it

The real state of the art isn't "pay the full tax every time" — it's pay for a lean high-frequency core up front, and the heavier orchestration tax only in sessions that reach for it.

Every tool you register is a standing tax on every call that agent's fleet makes. Five orchestration-shaped tools cost more than the entire core action layer combined. The design lesson is not "give agents fewer tools." It's separate the always-on core from the judgment-heavy periphery, and gate the periphery behind a discovery step.

Fig. 7 — Architecture for discipline that travels with capability

Builder checklist

Beautiful behavior is not free and it is not learned once. It is re-purchased, in tokens, every time the model looks at what it could do next. The report priced that purchase at roughly 18,000 tokens a session. What it's actually pricing is discipline.