Managing AI Agents: The New Playbook for High-Velocity Development

The Shift to Agent Management

Ryan Carson argues that the role of the knowledge worker has fundamentally changed: everyone is now a manager of AI agents. Success in this new paradigm requires moving away from local development environments toward cloud-based agent harnesses. By treating agents as a team rather than a single tool, developers can scale their output horizontally, managing multiple concurrent workstreams without the friction of local environment collisions or dependency management.

Cloud-First Development

Carson emphasizes that local development is becoming a bottleneck. By utilizing cloud VMs (via tools like Devin or similar agent harnesses), developers can spin up fresh, isolated environments for every task. This allows for parallel execution of 5-10 agents simultaneously. This architecture removes the mental overhead of context switching and synchronization, enabling a massive increase in shipping velocity—often 20 to 40 pull requests per day.

High-Stakes Decision Making

Managing agents at this scale requires a shift in cognitive load. Instead of writing code, the operator's primary job is making 10-20 high-stakes decisions before lunch. Carson advocates for a disciplined cadence—checking threads every 25 minutes—to avoid burnout. He maintains a hybrid workflow, often managing these threads from his phone, which allows him to keep agents moving even when away from his desk.

Automations and Self-Improvement

To maintain quality at high velocity, Carson relies on three core automations: end-to-end browser testing for signups, a daily production watchdog that summarizes customer activity, and a self-improvement loop where an agent (Grace) grades chat interactions against a rubric to suggest daily refinements. This creates a feedback loop where the system effectively improves its own performance over time.

Cost Management and Reputation

Token costs are managed through strict model routing, favoring cheaper, fine-tuned models for routine loop work while reserving top-tier models for complex tasks. Carson suggests a budget of roughly $5,000 per employee per month. Finally, he stresses the importance of building a public reputation by sharing these workflows on X, arguing that documenting the learning process is the most effective way to open doors and build long-term career durability.