AI Cascade & Agent Systems

The honest version of "AI agents in production" in 2026: most work is still cascades, not autonomous loops. A classifier decides whether the query is simple, a small fast model handles 80% of cases, a larger model handles the remaining 20%, and a human gets routed the 1% that need judgement. The agent isn't out-thinking you — it's plumbing that handles cost and quality together.

What I build

Cost-aware cascades — classification step routes to model tiers (Haiku → Sonnet → Opus, or Flash → Pro, or local-llama → frontier). Cascade decisions cached. Cost-per-query budgets enforced.

Tool-using agents — agent loops with a tool registry, retry semantics, max-step termination, and audit logs of every tool call. Built for use-cases that genuinely need multi-step reasoning, not bolted onto things that don't.

MCP servers — Model Context Protocol servers exposing your databases, APIs, or workflows as tools that any MCP-compatible AI client can use. Authenticated, rate-limited, observable.

Observability — every step traced with OpenTelemetry, LLM spans showing prompt/output/cost/latency, sampled traces drilled into Honeycomb / Langfuse. The "why is this slow?" question becomes a 5-minute investigation.

Operational discipline — eval suites covering the cascade end-to-end, regression detection on routing decisions, kill-switches and circuit breakers, gradual rollout via feature flags.

What I steer you away from

Wiring a 12-tool agent loop onto a problem that's a 200-line script. Multi-agent "swarm" architectures because someone read a paper. Autonomous agents that touch production data without human-in-the-loop checkpoints. Cascades fail less spectacularly than autonomous agents, and ship years sooner.