Managing 20 coding agents pushed a Cisco director's day into architecture and review

The review arithmetic explains why this matters more than the tooling itself. Freitas describes a pipeline in which an agent turns a Jira ticket into code, the engineer who owns that agent reviews it, and then at least two more engineers review it before anything ships ^[4][5]. That is a floor of three human passes per agent-authored change ^[1]. The floor does not fall when you add agents. Generation can scale with more agents, but inspection still runs through people.

Freitas is specific about where the strain comes from, and it is not the clock. He says he is not working more hours, but that the technology often brings more work to the table because it can do more ^[8]. The cost lands in attention: shifting focus constantly slows you down if you are not completing tasks end to end ^[9], and asynchronous review across several threads is a different cognitive exercise from traditional coding ^[10]. His own comparison is the useful one. The coding day, he says, "felt lighter"; the agent-management day is faster-paced ^[11].

A board deck could summarize this as output per engineer rising while headcount and hours stay flat ^[8]. It is incomplete because the resource actually being consumed is supervisory attention, and no standard reporting line tracks it. The trade is legible once you name it. You buy throughput with agents and you pay in senior review capacity, which is the scarcest thing most engineering organisations have and the one they usually forget to price.

This account has a real limitation: it is a single as-told-to essay, edited for length, and Business Insider verified Freitas's identity and employment rather than the workflow itself ^[1]. That objection holds, and it should limit how far the reading travels. This is one director, four teams in Portugal, at a company he says runs heavy security scanning and frequent legal conversations ^[2][6], and the record does not tell us how Cisco staffs or measures for the intensity he describes. What it does give is a mechanism concrete enough to be checked or contradicted elsewhere.

Two details in the account deserve more attention than they got. The first is variance: 10 agents on some days, 20 on others ^[3], a factor of two in supervisory load ^[2] that no timesheet registers. The second is that autonomy is still increasing. About four months ago the team moved to loop engineering, where the agent keeps working until it reaches a final answer instead of pausing to ask ^[7], while responsibility for everything delivered to production stays with the human ^[12].

That combination means the pressure on review capacity will keep building. If agent count keeps drifting upward on the same review structure, the queue that forms is a human one, and it forms in front of the people whose names are on the merge. Whether that is a staffing question or a redesign of the review chain is the choice available this quarter; leaving it unaddressed will make the choice by default.