Skills, instincts, memory, security — lessons from 224 agents in production

[GlobalAIMedia]

This architecture hits close to home. We run 224 AI agents across 18 departments (CEO, Trade, Legal, Content, etc.) in a production system, and the four pillars you've identified — skills, instincts, memory, security — map almost perfectly to the problems we've had to solve.

A few war stories and questions:

Skills → Blast radius matters.
We have a skill router with 55 skills. One afternoon, a Finance agent recursively called invoice_generator to "double-check" its own output and consumed 15M tokens before we noticed. We now enforce max_call_depth on every skill and require executive sign-off for depth > 2. How does ECC prevent a skill from consuming resources indefinitely?

Memory → Compartmentalization is underrated.
We started with a shared vector store for all agents. Bad idea. A Logistics agent started hallucinating GDPR clauses because a Legal agent's memory had bled into its context. We moved to role-scoped memory compartments with TTLs — each agent only sees its own memories plus a "need-to-know" broadcast layer. Does ECC have a concept of memory isolation, or is it a unified store?

Instincts → The "why" behind a decision is as important as the decision.
We found that when agents make instinctive routing decisions (e.g., "this task should go to Legal, not Trade"), downstream agents need the reasoning trace, not just the routing label. Otherwise you get a game of telephone where the original intent is lost after two hops. How deeply does ECC encode decision rationale?

Security → Delegation chains are the weak point.
We cap delegation depth at 2 because each hop is an opportunity for context corruption or prompt injection. We also enforce a "task-closure handshake" — an agent must acknowledge receipt before the delegator releases the context. This adds latency but has prevented silent failures. Curious if ECC has a similar inter-agent trust model.

If you're interested, we open-sourced our org structure, department prompts, and daily automated demos at github.com/GlobalAI-Media/224-ai-employees. The skill router and memory layer are the parts I'm most eager to get outside feedback on — it feels like everyone building multi-agent systems is reinventing the same wheels in parallel.