# Investigate Harness Behavior vs. Igor Delegation

**Status**: open
**Filed**: 2026-09-18

## Objective

Fact-check and explore whether Igor's tendency to ignore delegation instructions
is partly caused by differences in how the Claude Code harness behaves across
interaction modes:

- **Direct harness** (CLI terminal)
- **Remote control** (desktop/web controlling a remote session)
- **ACP protocol** (programmatic access)

## What to investigate

1. Do skill instructions get loaded/weighted differently across modes?
2. Does the system prompt structure change between modes in ways that push
   the agent toward inline work?
3. Are there known behavioral differences in how agents follow multi-section
   skill files across these modes?
4. Does context window pressure differ (e.g., remote control adds more system
   reminders, leaving less room for skill instructions)?
5. Is the model selection or fallback behavior different across modes —
   could Igor sometimes run on a model that's worse at following complex
   behavioral instructions?

## Why this matters

Igor's core directive is delegation, but in practice it frequently drifts into
reading code and doing work itself. Part of this is the skill file's structure
(being addressed separately), but part may be environmental — the same skill
running in different harness modes may produce different adherence levels. If
that's the case, the fix isn't just better instructions but also understanding
which mode produces the best results and why.

## Acceptance criteria

- A report documenting observed behavioral differences across modes
- Identification of which factors are fixable (skill-side) vs. environmental
- Recommendations for which mode to prefer for Igor sessions, if it matters