Plataforma SaaS para gestión y control de orquestación multi-agente con criterios verificables de finalización y presupuestos configurables.

Repository: rjmurillo/ai-agents Author: rjmurillo ## File `.claude/agents/orchestrator.md` ## Problem Line in Core Identity section: > **CRITICAL**: Only terminate when the problem is completely solved and ALL TODO items are checked off. "Completely solved" is not a verifiable condition. While the SESSION END GATE later in the file is detailed and good, this early directive could cause the orchestrator to loop indefinitely on subjective completion assessment before reaching the session end gate. ### Additional concern: No orchestration budget The orchestrator can spawn unlimited sub-agents with no mention of: - Maximum number of agent delegations per task - Token budget ceiling for the full orchestration - Time limit or iteration cap ## Why this matters (Claude 4.7) Claude 4.7's improved agentic capabilities make it more likely to take "completely solved" literally and over-invest in perfection. Without a budget, multi-agent orchestration can compound costs. ## Recommended changes 1. Replace vague stop criterion: ``` **CRITICAL**: Terminate when ALL TODO items are checked off AND the SESSION END GATE passes. If after 10 agent delegations the problem is not resolved, summarize progress, document remaining gaps, and stop. ``` 2. Add orchestration budget: ``` ## Orchestration Budget - Max agent delegations per task: 15 (warn at 10) - If budget exhausted: summarize progress, list remaining work, return to user ```