Managing over 10 agents at Cisco does not extend my hours but makes the work feel more demanding.

As told to

This as-told-to essay is based on a conversation with Sergio Freitas, senior director of engineering at Cisco, based in Lisbon, Portugal. His identity and employment have been verified by Business Insider. This story has been edited for length and clarity.

I began my career as a software engineer, advanced into management, and now serve as a director of engineering at Cisco, leading four teams located in Portugal.

Software engineering has always involved ongoing evolution. New tools such as programming languages or frameworks require constant learning. Recently the core methods of how engineers operate have shifted quickly, especially over the past year, creating adaptation challenges and added pressure.

The major shift occurred after advanced models became available late last year, including Anthropic's Claude Opus 4.5. At that point AI became deeply integrated into daily processes, enabling us to assign work and collaborate with agents on engineering tasks. Prior use of AI was limited mainly to suggestions and basic prompts.

Agentic AI transforms engineering roles

Daily responsibilities have moved from writing code directly to emphasizing architecture, design, workflow oversight, and agent coordination. Engineers now act more as system architects and process managers, tackling bigger and more intricate design challenges.

The agents function like team members. I oversee 10 to 20 agents each day that in turn direct other agents. The count varies with task complexity. A primary coordinating agent can launch supporting agents for targeted work.

These agents operate continuously and monitor updates from collaboration tools, GitHub requests, and other connected systems when authorized. For instance, an agent can generate code from a Jira ticket description, after which the assigned engineer reviews it. The process then requires approval from at least two engineers.

Four months ago we adopted loop engineering, allowing agents to continue iterations independently until reaching a complete result rather than pausing for questions.

At times agents request permission to access specific systems or execute operations. Autonomy is increasing yet remains bounded by clear controls.

Reviewing expanded output

AI collaboration offers clear benefits, yet engineers must stay attentive during code reviews. All production releases remain the responsibility of humans rather than agents.

My total working hours have not increased, but AI often expands the volume of deliverables because it enables greater output.

Cisco places strong emphasis on security, requiring multiple scans and discussions with legal teams.

A quicker rhythm emerges

Engineers adjusting to this approach may experience initial strain from handling higher volumes. Frequent context switching can reduce efficiency when tasks are not finished sequentially.

Adapting to elevated daily output and parallel work with agents requires new routines for handling multiple streams and asynchronous checks, which differ from conventional coding focus.

The overall scope has broadened. By day's end the effort feels more concentrated, though not through the prolonged immersion once needed for coding.

Writing code was once a specialized ability that felt empowering, like building something original. Engineers who value that hands-on creation may find the transition difficult.

Previously the close of the day often felt lighter. Managing agents now creates a quicker tempo.