How small senior teams use AI agents without surrendering judgment
There are two fashionable positions on AI coding agents: that they replace engineering teams, and that they’re a liability serious teams should avoid. We hold neither. We run agents in our own production delivery every day, and the honest lesson is narrower than both camps want it to be: agents multiply a senior team’s output, and they do it safely only when the humans keep specific, non-negotiable gates.
Why this matters most for small teams
A large organization can afford to treat agents as a curiosity. A small senior team can’t — the leverage is too real. Work that used to force a choice between hiring and declining now gets done by agents operating in parallel, with the senior engineers spending their hours where judgment actually lives: deciding what to build, reviewing what was built, and owning what ships. Delivery capacity stops being a headcount question. That’s not a prediction; it’s how our own products and client work get built, daily.
The catch is that leverage and risk scale together. An agent that can write real code against a real repository can also mangle one. The teams that get the multiplier without the incidents are the ones that treat judgment as an architecture problem, not a vibe.
Three decisions that never leave human hands
Everything between these gates can be agent work. The gates themselves cannot.
-
What gets built
Scope is signed off by a human before an agent touches a task. Agents are excellent at doing the thing; they are unreliable at asking whether the thing is worth doing. A queue where work is scoped and claimed atomically — rather than improvised in a chat window — is what keeps a fleet pointed at the right problems.
-
What ships
Nothing an agent writes reaches production without human review and merge. This is the same discipline any senior team applies to any contributor — the agent is a prolific junior colleague with unusual failure modes, not an exempt one. Review is where generated code’s quiet corner-cutting gets caught, and it’s cheap compared to the incident it prevents.
-
What the fleet may touch
Agents never hold real credentials. Each holds a scoped, revocable token; anything sensitive is mediated by systems that enforce policy at the transport layer, where a confused model can’t route around it. Prompts are advisory. Infrastructure is not. If your agent safety story is a paragraph in a system prompt, you don’t have one.
This is old discipline in new clothes
None of this was invented for LLMs. We shipped AI in production in 2018 — Speak Meetings, custom NLP pipelines running on live call traffic years before large language models were a commodity — and the habits from that era transfer directly: treat the model as a component, evaluate before trusting, expect failure as a distribution rather than an exception. What changed with agents is the blast radius, which is why the gates moved from good practice to load-bearing.
Our founder writes about the operating detail — the work queue as system of record, credential custody, transport-enforced git policy — in Running a fleet of AI coding agents in production. The short version: every part of the discipline exists because something went wrong without it.
Two ways this shows up in our work
First, indirectly: when you hire us to rescue a prototype or build an operational system, the fleet’s leverage shows up as your finished product — senior judgment at the gates, agent throughput between them. It’s also why we can say from daily practice, not posture, what AI-generated code is worth — the theme of when a vibe-coded prototype needs senior engineering.
Second, directly: if your own team wants agents doing real work and is rightly unwilling to hand them credentials and hope, building the gates is itself an engagement — AI agent infrastructure: workflows, credential mediation, audit trails, and approval gates that are enforced rather than requested.
Want agents doing real work, safely?
Joseph reads every note and replies himself. If agents aren’t the right tool for your problem yet, he’ll say so.
Start a Project