Past a certain number of agentic systems in production, decisions about acceptable risk, tooling standards, and cross-team consistency need a body that isn't any single product or platform team. A governance council is that body, done well or done as theater."

The organizations managing AI costs well treat it as a dedicated discipline with real ownership, not a line item someone notices when the invoice arrives. Here's what that discipline — TokenOps — actually looks like.

The earlier platform-team post covered when and what to build. This picks up the org-design question specifically — reporting lines, funding model, and the structural choices that determine whether a platform team actually gets adopted or gets quietly routed around."

The earlier SLA post covered external commitments to other teams. SLOs — the internal target a team holds itself to — need to exist even before, or instead of, a formal external SLA, and retrofitting them onto a system with no baseline is its own specific challenge."

The fallback chain from the previous post handles outages. True multi-provider redundancy — actively split across providers, not just failover — raises a different question: how to get the reliability benefit without paying for full duplicate capacity at all times.

An LLM provider outage is outside your control entirely, and a single-provider architecture means that outage becomes your outage too. A fallback chain is the direct mitigation, with its own real design tradeoffs."

Routing simple requests to a cheap model and complex ones to a frontier model is the idea everyone agrees with. The actual routing decision logic — how "complex" gets determined reliably and cheaply — is where most of the real engineering effort goes."

Switching to a cheaper model and watching the error rate is necessary and not sufficient — quality degradation from a downgrade is often subtle enough that error-rate monitoring alone won't catch it.

A false fact planted in an agent's long-term memory doesn't go away when you correct it in conversation — it's still sitting in the store, waiting to influence the next session. Here's how to actually defend against it.

For a narrow, high-volume task, a distilled small model trained on a frontier model's outputs can match quality at a fraction of the cost — genuinely one of the highest-leverage cost optimizations available, and one of the most misapplied when used outside its actual sweet spot."