Summary
Agent harness context economics is the practice of accounting for the per-turn token cost of a coding agent's extensibility surfaces—instructions, skills, tools, subagents, and hooks—and attributing context bloat to the configuration that causes it. Anthropic documents these costs for Claude Code; the same accounting logic applies to any harness with similar extensibility tiers. The core warning is that a large always-on instruction file, excessive tool discovery, or redundant parallel subagents quietly taxes every interaction, and teams often blame the harness for bloat their own configuration created.
Key Points
- Extensibility surfaces have different cost tiers: some enter every request, some load only on use, and some cost nothing unless they produce output.
- A giant
CLAUDE.mdtaxes every turn; too many MCP tools expand discovery overhead; too many parallel subagents duplicate investigation. - To avoid misattributing bloat, run a minimal configuration against the proposed production configuration in a pilot.
- Hooks are the cheapest surface: they consume zero model tokens unless they return output.
- Claude Code v2.1.224 (2026-08-07) adds cross-session messaging; sessions coordinate as peers rather than a parent-child fan-out tree.
Concepts
- Always-on tier: content injected into every request (e.g.,
CLAUDE.md). - Two-tier skill model: skill descriptions are always present; full skill bodies load only when used.
- On-demand tool schemas: MCP tool schemas can be loaded when needed instead of up front.
- Separate subagent context: subagents carry their own context window, isolating their cost from the main thread.
- Zero-token hooks: hooks only consume tokens if they return output.
- Attribution pilot: comparing minimal vs. production configurations to determine which layer causes context bloat.
- : Aider uses a graph-ranked repository map with an explicit configurable budget defaulting to ~1000 tokens, expanded on demand.