AI platforms went down 51 times in Q1. Our agents kept running.

We learned this the hard way when our agents went dark for 6 hours during the Claude outage in March. Our support agent stopped responding to tickets. Our coding agent abandoned a deployment mid-push. Our social media agent missed 14 scheduled posts.

The problem wasn't just downtime — it was silent failure. No alerts. No graceful degradation. Just agents that looked like they were working while doing absolutely nothing.

Here's the redundancy stack that keeps our agents running when platforms go down:

Model Failover Chain: Primary model → Secondary model → Cached response → Human escalation

Your agent config needs a model hierarchy, not just a model preference:

model_chain:
  primary: "claude-3-5-sonnet-20241022"
  secondary: "gpt-4o"
  fallback: "gemini-pro"
  cache_duration: "24h"
  escalation_threshold: 3

But model switching isn't enough. You need capability-aware failover — different models handle different tasks differently. Our coding agent routes complex refactoring to Claude but falls back to GPT-4o for simple fixes. Our support agent uses Claude for nuanced responses but switches to GPT-4o for straightforward FAQ answers.

The real lesson: test your failover under load. We thought our secondary models could handle the traffic. They couldn't. When Claude went down, our GPT-4o costs spiked 340% and we hit rate limits within 20 minutes.

Now we pre-negotiate rate limit increases with backup providers and run monthly failover drills. Every first Friday, we kill our primary model connection and let agents run on backup infrastructure for 2 hours.

Critical: Your agents need to detect platform degradation, not just outages. Slow responses and high error rates kill productivity before total outages do.

We built health checks that monitor response time and success rates:

health_check:
  response_time_threshold: "5s"
  error_rate_threshold: "10%"
  check_interval: "30s"
  failover_after: "3_consecutive_failures"

The pattern that changed everything: graceful degradation with capability mapping. Instead of all-or-nothing failover, we map tasks to model capabilities. Complex reasoning stays on premium models. Simple tasks get routed to cheaper, more reliable alternatives immediately.

Document your failover scenarios now. When platforms go down, you won't have time to figure out which model handles which task type. Your agents will just stop working, and you'll spend the outage debugging instead of serving customers.