Agents
Production AI Agents
Most "AI agents" are chatbots that answer questions. The useful ones change state: they create the shipment, place the order, resolve the exception. That is a different engineering problem, because now being wrong has consequences. We build agents that write, with the guardrails that makes safe.
- Vercel AI SDK
- Anthropic
- OpenAI
- AI Gateway
- Zod
- MCP
What this includes
Each of these exists because leaving it out is how these systems fail in production.
Tool design
Each action your agent can take becomes an explicit, typed, individually testable tool. Schemas are enforced at the boundary, so a malformed model response fails loudly instead of writing bad data.
Approval gates
Write actions are gated server-side, not by prompt instruction. High-consequence operations are built read-only by design: the agent proposes, an operator applies. A model cannot talk its way past a gate that lives in your backend.
Idempotency and recovery
Retries, partial failures, and duplicate submissions are handled explicitly. An agent that runs twice does not create two orders.
Multi-provider routing
Models are configured, not hardcoded, with per-provider fallback ordering. When a provider degrades or prices change, you switch without a rewrite — and you are never locked to one vendor.
13
Production tools in one shipping agent
7
Models with per-provider fallback ordering
Server
Where approval gates are enforced
Next step
Tell us what you’re trying to build
Most engagements start with a fixed-price audit, so the first thing you buy is a decision rather than a commitment.