AI Agent Execution Constraints

Overview

Operational constraints for AI agent execution: loop budgets (10-20 iterations), spec-first philosophy, context boundary enforcement, and prohibition of manual edits during automated execution. Provides a general framework for constraining agentic behavior beyond tool-specific configuration.

Constraints

Loop Budgets

• Soft limit: 10 iterations for standard tasks
• Hard limit: 20 iterations before mandatory human review
• Purpose: Prevent runaway loops and wasted compute

Spec-First Philosophy

• Agents must operate from an explicit specification
• No ad-hoc code generation without a governing spec
• Spec violations trigger automatic halt

Context Boundaries

• Agents see only their designated field of vision
• Sensitive paths are blacklisted via manifest/allowlist
• Spec and planning paths are read-only

Execution Discipline

• No manual edits during automated execution
• Commit after every meaningful edit
• Mandatory approval for destructive operations

See Also

• agent-scratchpad-workspace - Workspace isolation pattern for agents
• gemini-loop-prevention - Tool-specific loop prevention for Gemini CLI
• human-in-the-loop-approval-gates - Approval gate taxonomy for agentic safety