Agent Ops: how to stop being the human API

As agentic systems become widespread, I'm starting to see something that could be called Agent Ops. It means designing automation architecture that solves the biggest problem out there, which is that we humans will always be the bottleneck in the automation cycle.

The highest-priority task is deciding which communications can bypass human intervention. Picture an operations worker who handles requests coming in through Slack every day and finds themselves acting as a human API. That API connects all the incoming messages with the internal knowledge base, call recordings, Google Workspace and internal tools.

For example, where they once checked and answered emails all day, they now grab transcripts of client and customer calls and turn them into prompts. Then they feed those prompts into AI and send the output back.

I usually sit down with pencil and paper and diagram all the workflows and how they can split into different streams. Once the streams are formalised, I create one agent per stream, each with its own one-pager of documentation that humans read. Every agent also gets its own Asana/Jira ticket space and a memory thread where it offloads everything that passes through it each day. Basically, it's like the microservices architecture of the old days, where everything is a stream. That's difficult, but it splits your problems into different buckets and lets you focus on architecture separately from the content of the actual messages.

Surprisingly, working on automation architecture doesn't really give me more free time, as my work just shifts toward wider decision-making. For example, I can now focus on skill design, writing evals and designing better memory systems.

Yours, one link deep,
Zhenia