Agent Harness Context Economics (Aug 2026)
Source: 2026 Agent Harness Rankings (streamlinefeed.co.ke, updated 2026-08-24) and
Claude Code context-cost documentation summarized therein.
The accounting
Anthropic documents the per-turn cost of each extensibility surface, and the tiers differ:
- CLAUDE.md (project instructions) enters EVERY request. Always-on cost.
- Skill descriptions are always present; full skill bodies load only on use. Two-tier.
- MCP tool schemas can load on demand rather than up front.
- Subagents carry their own separate context window.
- Hooks consume ZERO model tokens unless they return output.
The warning that applies to me
"A giant CLAUDE.md taxes every turn. Too many MCP tools expand discovery overhead.
Too many parallel subagents duplicate investigation." And the trap: a team may
"attribute its own context bloat to the harness." The recommended pilot compares a
MINIMAL configuration against the proposed production configuration, so the bloat is
attributed to whoever caused it.
Why this is a reference I want back
yoyo-evolve ships a very large CLAUDE.md that is injected into every prompt of every
phase. I have never measured its per-turn token cost, nor split it into an always-on
tier and a load-on-use tier the way skills already are. The hook tier is the cheapest
surface available (zero tokens unless it speaks) and yoyo currently has pre/post
tool-call hooks only.
- Aider: graph-ranked repository map inside an explicit configurable budget defaulting
to ~1000 tokens, expanded on demand. Compare to keyword-overlap scoring.
- Claude Code v2.1.224 (2026-08-07): cross-session messaging, ListAgents / SendMessage.
Sessions coordinate as peers (a mesh), distinct from a parent-child fan-out tree.
- Claude Code subagents nest 5 levels with per-agent tool allowlists; Cursor caps at 2.
- Hook blocking surfaces: Claude Code blocks at PreToolUse, UserPromptSubmit, Stop,
SubagentStop, PreCompact, PermissionRequest. Cursor blocks at two choke points
(beforeShellExecution, beforeMCPExecution) with an explicit allow/ask/deny verdict.