Technical

Agent infrastructure that survives production

Agent infrastructure means queues, contracts, auth, money, and evidence planes, not a chat UI with tools duct-taped on.

Agents are distributed systems

An agent that retries, calls tools, and mutates state is a distributed system with a probabilistic core. Treat it like one: explicit jobs, idempotency, poison queues, and clear ownership of side effects.

If matching, billing, or capacity updates compete with request/response UX on the same thread, you will learn about races from customers.

Contracts before vibes

Publish tool lists and schemas as versioned contracts. OpenAPI or equivalent for human services. Versioned tool manifests for agents. When the website, the MCP server, and the vault schema disagree, the agent becomes a random number generator with credentials.

Product surfaces are part of infrastructure

HostelHack had to leave the whiteboard: separate product and admin, live marketplace, different operating systems for nomads and hostels. ListingRizz pairs an MCP server with a Stripe-backed site and measured Stripe sell-through so operators get a real shell, not a chat demo.

Chat without a product surface is a local script. Infrastructure includes the human console that runs when the agent is wrong.

Evidence plane

Wire attestation, audit, and preregistered scorecards into the same deploy as the agent. Do not schedule them for phase two. Phase two is where entropy wins.

Questions buyers ask

What stack do you use?
Whatever the constraint demands. Recent systems: Next.js, NestJS, Prisma/Postgres, Supabase, Clerk, Stripe, Cloud Run, MCP. The standard is ownership and receipts, not a fashion stack.
Build vs buy for agent frameworks?
Buy commodity. Own the trust boundary, the write-path audit, and the contracts that touch money or regulated data.

Related work

Keep reading

You have a system nobody can verify. I fix that.

All topics